[BugFix]: fix a lot of bug (#1565)

Signed-off-by: princepride <wangzhipeng628@gmail.com>

.buildkite/pipeline.yml

@@ -3,347 +3,38 @@ steps:
     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 --file docker/Dockerfile.ci -t vllm-omni-ci ."
+      - "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"
 
-  # - label: "Test on NPU"
-  #   depends_on: ~
-  #   key: npu-test
-  #   commands:
-  #     - ".buildkite/scripts/hardware_ci/run_npu_test.sh"
-  #   agents:
-  #     queue: "ascend"
-
-  - label: "Simple Unit Test"
+  # L2 Test
+  - label: "Upload Ready Pipeline"
     depends_on: image-build
+    key: upload-ready-pipeline
+    if: build.branch != "main"
     commands:
-    - |
-      pytest -v -s \
-        tests/entrypoints/ \
-        tests/diffusion/cache/ \
-        tests/diffusion/lora/ \
-        tests/model_executor/models/qwen2_5_omni/test_audio_length.py \
-        tests/worker/ \
-        tests/distributed/omni_connectors/test_kv_flow.py \
-        --cov=vllm_omni \
-        --cov-branch \
-        --cov-report=term-missing \
-        --cov-report=html \
-        --cov-report=xml
+      - buildkite-agent pipeline upload .buildkite/test-ready.yml
     agents:
-      queue: "gpu_1_queue"
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+      queue: "cpu_queue_premerge"
 
-  - label: "Diffusion Model Test"
-    timeout_in_minutes: 20
+  # L3 Test
+  - label: "Upload Merge Pipeline"
     depends_on: image-build
+    key: upload-merge-pipeline
+    if: build.branch == "main"
     commands:
-      - pytest -s -v tests/e2e/offline_inference/test_t2i_model.py
+      - buildkite-agent pipeline upload .buildkite/test-merge.yml
     agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+      queue: "cpu_queue_premerge"
 
-  - label: "Diffusion Images API LoRA E2E"
-    timeout_in_minutes: 20
+  # L4 Test
+  - label: "Upload Nightly Pipeline"
     depends_on: image-build
+    key: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1"
     commands:
-      - pytest -s -v tests/e2e/online_serving/test_images_generations_lora.py
+      - buildkite-agent pipeline upload .buildkite/test-nightly.yml
     agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion Model CPU offloading Test"
-    timeout_in_minutes: 20
-    depends_on: image-build
-    commands:
-      - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
-      - pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
-    agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Audio Generation Model Test"
-    timeout_in_minutes: 20
-    depends_on: image-build
-    commands:
-      - pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
-    agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion Cache Backend Test"
-    timeout_in_minutes: 15
-    depends_on: image-build
-    commands:
-      - pytest -s -v tests/e2e/offline_inference/test_cache_dit.py tests/e2e/offline_inference/test_teacache.py
-    agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion Sequence Parallelism Test"
-    timeout_in_minutes: 20
-    depends_on: image-build
-    commands:
-      - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion Tensor Parallelism Test"
-    timeout_in_minutes: 20
-    depends_on: image-build
-    commands:
-      - pytest -s -v tests/e2e/offline_inference/test_zimage_tensor_parallel.py
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion GPU Worker Test"
-    timeout_in_minutes: 20
-    depends_on: image-build
-    commands:
-      - pytest -s -v tests/diffusion/test_diffusion_worker.py
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-
-  - label: "Benchmark Test"
-    timeout_in_minutes: 15
-    depends_on: image-build
-    commands:
-      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - pytest -s -v tests/benchmarks/test_serve_cli.py
-    agents:
-      queue: "mithril-h100-pool"
-    plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                resources:
-                  limits:
-                    nvidia.com/gpu: 2
-                volumeMounts:
-                  - name: devshm
-                    mountPath: /dev/shm
-                  - name: hf-cache
-                    mountPath: /root/.cache/huggingface
-                env:
-                  - name: HF_HOME
-                    value: /root/.cache/huggingface
-            nodeSelector:
-              node.kubernetes.io/instance-type: gpu-h100-sxm
-            volumes:
-              - name: devshm
-                emptyDir:
-                  medium: Memory
-              - name: hf-cache
-                hostPath:
-                  path: /mnt/hf-cache
-                  type: DirectoryOrCreate
-
-  - label: "Omni Model Test"
-    timeout_in_minutes: 15
-    depends_on: image-build
-    commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-    - docker#v5.2.0:
-        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-        always-pull: true
-        propagate-environment: true
-        environment:
-        - "HF_HOME=/fsx/hf_cache"
-        volumes:
-        - "/fsx/hf_cache:/fsx/hf_cache"
-
-  # - label: "Omni Model Test with H100"
-  #   timeout_in_minutes: 30
-  #   depends_on: image-build
-  #   commands:
-  #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  #     - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
-  #     - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
-  #     - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py
-  #     - pytest -s -v tests/e2e/online_serving/test_async_omni.py
-  #   agents:
-  #     queue: "mithril-h100-pool"
-  #   plugins:
-  #     - kubernetes:
-  #         podSpec:
-  #           containers:
-  #             - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-  #               resources:
-  #                 limits:
-  #                   nvidia.com/gpu: 2
-  #               volumeMounts:
-  #                 - name: devshm
-  #                   mountPath: /dev/shm
-  #                 - name: hf-cache
-  #                   mountPath: /root/.cache/huggingface
-  #               env:
-  #                 - name: HF_HOME
-  #                   value: /root/.cache/huggingface
-  #           nodeSelector:
-  #             node.kubernetes.io/instance-type: gpu-h100-sxm
-  #           volumes:
-  #             - name: devshm
-  #               emptyDir:
-  #                 medium: Memory
-  #             - name: hf-cache
-  #               hostPath:
-  #                 path: /mnt/hf-cache
-  #                 type: DirectoryOrCreate
-
-  - label: "Diffusion Image Edit Test with H100 (1 GPU)"
-    timeout_in_minutes: 20
-    depends_on: image-build
-    commands:
-      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
-    agents:
-      queue: "mithril-h100-pool"
-    plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                resources:
-                  limits:
-                    nvidia.com/gpu: 1
-                volumeMounts:
-                  - name: devshm
-                    mountPath: /dev/shm
-                  - name: hf-cache
-                    mountPath: /root/.cache/huggingface
-                env:
-                  - name: HF_HOME
-                    value: /root/.cache/huggingface
-            nodeSelector:
-              node.kubernetes.io/instance-type: gpu-h100-sxm
-            volumes:
-              - name: devshm
-                emptyDir:
-                  medium: Memory
-              - name: hf-cache
-                hostPath:
-                  path: /mnt/hf-cache
-                  type: DirectoryOrCreate
-
-  # - label: "Bagel Text2Img Model Test with H100"
-  #   timeout_in_minutes: 30
-  #   depends_on: image-build
-  #   commands:
-  #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  #     - pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py
-  #   agents:
-  #     queue: "mithril-h100-pool"
-  #   plugins:
-  #     - kubernetes:
-  #         podSpec:
-  #           containers:
-  #             - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-  #               resources:
-  #                 limits:
-  #                   nvidia.com/gpu: 1
-  #               volumeMounts:
-  #                 - name: devshm
-  #                   mountPath: /dev/shm
-  #                 - name: hf-cache
-  #                   mountPath: /root/.cache/huggingface
-  #               env:
-  #                 - name: HF_HOME
-  #                   value: /root/.cache/huggingface
-  #           nodeSelector:
-  #             node.kubernetes.io/instance-type: gpu-h100-sxm
-  #           volumes:
-  #             - name: devshm
-  #               emptyDir:
-  #                 medium: Memory
-  #             - name: hf-cache
-  #               hostPath:
-  #                 path: /mnt/hf-cache
-  #                 type: DirectoryOrCreate
+      queue: "cpu_queue_premerge"

.buildkite/test-amd.yaml

@@ -33,6 +33,7 @@ steps:
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+    - pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
 
 - label: "Diffusion Cache Backend Test"
   timeout_in_minutes: 15
@@ -44,7 +45,7 @@ steps:
     - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_cache_dit.py tests/e2e/offline_inference/test_teacache.py
+    - pytest -s -v -m 'core_model and cache and diffusion and not distributed_rocm and MI325'
 
 - label: "Diffusion Sequence Parallelism Test"
   timeout_in_minutes: 20
@@ -56,7 +57,7 @@ steps:
     - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
+    - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py -m core_model
 
 - label: "Diffusion Tensor Parallelism Test"
   timeout_in_minutes: 20
@@ -66,7 +67,7 @@ steps:
     - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_zimage_tensor_parallel.py
+    - pytest -s -v tests/e2e/offline_inference/test_zimage_parallelism.py
 
 - label: "Diffusion GPU Worker Test"
   timeout_in_minutes: 20
@@ -88,6 +89,7 @@ steps:
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    - pytest -s -v tests/engine/test_async_omni_engine_abort.py
 
 - label: "Omni Model Test Qwen3-Omni"
   timeout_in_minutes: 15
@@ -101,7 +103,7 @@ steps:
     - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
     - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
     - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py
-    - pytest -s -v tests/e2e/online_serving/test_async_omni.py
+
 
 - label: "Diffusion Image Edit Test"
   timeout_in_minutes: 15

.buildkite/test-merge.yml

@@ -0,0 +1,308 @@
+steps:
+  - label: "Simple Unit Test"
+    depends_on: upload-merge-pipeline
+    commands:
+      - "pytest -v -s -m 'core_model and cpu' --cov=vllm_omni --cov-branch --cov-report=term-missing --cov-report=html --cov-report=xml"
+    agents:
+      queue: "gpu_1_queue"
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Model Test"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_t2i_model.py -m "advanced_model and diffusion" --run-level "advanced_model"
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Images API LoRA E2E"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/e2e/online_serving/test_images_generations_lora.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Model CPU offloading Test"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+      - pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Audio Generation Model Test"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Cache Backend Test"
+    timeout_in_minutes: 15
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v -m 'core_model and cache and diffusion and not distributed_cuda and L4'
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Sequence Parallelism Test"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Tensor Parallelism Test"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_zimage_parallelism.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion GPU Worker Test"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - pytest -s -v tests/diffusion/test_diffusion_worker.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Benchmark & Engine Test"
+    depends_on: upload-merge-pipeline
+    commands:
+      - |
+        timeout 15m bash -c '
+                export VLLM_WORKER_MULTIPROC_METHOD=spawn
+                set +e
+                pytest -s -v tests/benchmarks/test_serve_cli.py
+                EXIT1=$$?
+                pytest -s -v tests/engine/test_async_omni_engine_abort.py
+                EXIT2=$$?
+                exit $$((EXIT1 | EXIT2))
+        '
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Omni Model Test"
+    timeout_in_minutes: 15
+    depends_on: upload-merge-pipeline
+    commands:
+      - export VLLM_LOGGING_LEVEL=DEBUG
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Omni Model Test with H100"
+    timeout_in_minutes: 30
+    depends_on: upload-merge-pipeline
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
+      - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
+      - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "advanced_model" --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:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
+  - label: "Diffusion Image Edit Test with H100 (1 GPU)"
+    timeout_in_minutes: 20
+    depends_on: upload-merge-pipeline
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 1
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
+  # - label: "Bagel Text2Img Model Test with H100"
+  #   timeout_in_minutes: 30
+  #   depends_on: upload-merge-pipeline
+  #   commands:
+  #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  #     - pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py
+  #   agents:
+  #     queue: "mithril-h100-pool"
+  #   plugins:
+  #     - kubernetes:
+  #         podSpec:
+  #           containers:
+  #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+  #               resources:
+  #                 limits:
+  #                   nvidia.com/gpu: 1
+  #               volumeMounts:
+  #                 - name: devshm
+  #                   mountPath: /dev/shm
+  #                 - name: hf-cache
+  #                   mountPath: /root/.cache/huggingface
+  #               env:
+  #                 - name: HF_HOME
+  #                   value: /root/.cache/huggingface
+  #           nodeSelector:
+  #             node.kubernetes.io/instance-type: gpu-h100-sxm
+  #           volumes:
+  #             - name: devshm
+  #               emptyDir:
+  #                 medium: Memory
+  #             - name: hf-cache
+  #               hostPath:
+  #                 path: /mnt/hf-cache
+  #                 type: DirectoryOrCreate

.buildkite/test-nightly.yaml

@@ -1,45 +0,0 @@
-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 --file docker/Dockerfile.ci -t vllm-omni-ci ."
-      - "docker tag vllm-omni-ci public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT"
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT"
-    agents:
-      queue: "cpu_queue_premerge"
-
-  - label: "Omni Model Test with H100"
-    timeout_in_minutes: 180
-    depends_on: image-build
-    commands:
-      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_expansion.py
-    agents:
-      queue: "mithril-h100-pool"
-    plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                resources:
-                  limits:
-                    nvidia.com/gpu: 2
-                volumeMounts:
-                  - name: devshm
-                    mountPath: /dev/shm
-                  - name: hf-cache
-                    mountPath: /root/.cache/huggingface
-                env:
-                  - name: HF_HOME
-                    value: /root/.cache/huggingface
-            nodeSelector:
-              node.kubernetes.io/instance-type: gpu-h100-sxm
-            volumes:
-              - name: devshm
-                emptyDir:
-                  medium: Memory
-              - name: hf-cache
-                hostPath:
-                  path: /mnt/hf-cache
-                  type: DirectoryOrCreate

.buildkite/test-nightly.yml

@@ -0,0 +1,98 @@
+steps:
+  - label: "Omni Model Test with H100"
+    timeout_in_minutes: 90
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1"
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - |
+        set +e
+        pytest -s -v tests/e2e/online_serving/test_qwen3_omni_expansion.py -m "advanced_model" --run-level "advanced_model"
+        EXIT1=$$?
+        pytest -s -v tests/examples/online_serving/test_qwen3_omni.py -m "advanced_model" --run-level "advanced_model"
+        EXIT2=$$?
+        exit $$((EXIT1 | EXIT2))
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
+  - label: "Omni Model Test"
+    timeout_in_minutes: 60
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1"
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/examples/online_serving/test_qwen2_5_omni.py -m "advanced_model" --run-level "advanced_model"
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          shm-size: "8gb"
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Omni Model Perf Test"
+    timeout_in_minutes: 120
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1"
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/perf/scripts/run_benchmark.py
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate

.buildkite/test-ready.yml

@@ -0,0 +1,306 @@
+steps:
+  - label: "Simple Unit Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - "timeout 20m pytest -v -s -m 'core_model and cpu' --cov=vllm_omni --cov-branch --cov-report=term-missing --cov-report=html --cov-report=xml"
+    agents:
+      queue: "gpu_1_queue"
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Model Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - timeout 20m pytest -s -v tests/e2e/offline_inference/test_t2i_model.py -m "core_model and diffusion" --run-level "core_model"
+    agents:
+      queue: "gpu_1_queue"
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Model CPU offloading Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - |
+        timeout 20m bash -c '
+          set +e
+          pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+          EXIT1=$$?
+          pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+          EXIT2=$$?
+          exit $$((EXIT1 | EXIT2))
+        '
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Audio Generation Model Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Cache Backend Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - timeout 15m pytest -s -v -m 'core_model and cache and diffusion and not distributed_cuda and L4'
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Sequence Parallelism Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - timeout 20m pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py -m core_model
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion GPU Worker Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+
+  - label: "Benchmark & Engine Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - |
+        timeout 15m bash -c '
+                export VLLM_WORKER_MULTIPROC_METHOD=spawn
+                set +e
+                pytest -s -v tests/benchmarks/test_serve_cli.py
+                EXIT1=$$?
+                pytest -s -v tests/engine/test_async_omni_engine_abort.py
+                EXIT2=$$?
+                exit $$((EXIT1 | EXIT2))
+        '
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+
+  - label: "Omni Model Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - |
+        timeout 17m bash -c '
+          export VLLM_LOGGING_LEVEL=DEBUG
+          export VLLM_WORKER_MULTIPROC_METHOD=spawn
+          pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+        '
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  # - label: "Omni Model Test with H100"
+  #   depends_on: upload-ready-pipeline
+  #   commands:
+  #     - |
+  #       timeout 20m bash -c '
+  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  #         export VLLM_TEST_CLEAN_GPU_MEMORY="1"
+  #         # - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
+  #         pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
+  #       '
+  #   agents:
+  #     queue: "mithril-h100-pool"
+  #   plugins:
+  #     - kubernetes:
+  #         podSpec:
+  #           containers:
+  #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+  #               resources:
+  #                 limits:
+  #                   nvidia.com/gpu: 2
+  #               volumeMounts:
+  #                 - name: devshm
+  #                   mountPath: /dev/shm
+  #                 - name: hf-cache
+  #                   mountPath: /root/.cache/huggingface
+  #               env:
+  #                 - name: HF_HOME
+  #                   value: /root/.cache/huggingface
+  #           nodeSelector:
+  #             node.kubernetes.io/instance-type: gpu-h100-sxm
+  #           volumes:
+  #             - name: devshm
+  #               emptyDir:
+  #                 medium: Memory
+  #             - name: hf-cache
+  #               hostPath:
+  #                 path: /mnt/hf-cache
+  #                 type: DirectoryOrCreate
+
+  - label: "Qwen3-TTS E2E Test"
+    depends_on: upload-ready-pipeline
+    commands:
+      - |
+        timeout 20m bash -c '
+          export VLLM_LOGGING_LEVEL=DEBUG
+          export VLLM_WORKER_MULTIPROC_METHOD=spawn
+          pytest -s -v tests/e2e/online_serving/test_qwen3_tts.py
+        '
+    agents:
+      queue: "gpu_4_queue"
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  # - label: "Diffusion Image Edit Test with H100 (1 GPU)"
+  #   depends_on: upload-ready-pipeline
+  #   commands:
+  #     - |
+  #       timeout 20m bash -c '
+  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  #         pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
+  #       '
+  #   agents:
+  #     queue: "mithril-h100-pool"
+  #   plugins:
+  #     - kubernetes:
+  #         podSpec:
+  #           containers:
+  #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+  #               resources:
+  #                 limits:
+  #                   nvidia.com/gpu: 1
+  #               volumeMounts:
+  #                 - name: devshm
+  #                   mountPath: /dev/shm
+  #                 - name: hf-cache
+  #                   mountPath: /root/.cache/huggingface
+  #               env:
+  #                 - name: HF_HOME
+  #                   value: /root/.cache/huggingface
+  #           nodeSelector:
+  #             node.kubernetes.io/instance-type: gpu-h100-sxm
+  #           volumes:
+  #             - name: devshm
+  #               emptyDir:
+  #                 medium: Memory
+  #             - name: hf-cache
+  #               hostPath:
+  #                 path: /mnt/hf-cache
+  #                 type: DirectoryOrCreate
+
+  # - label: "Bagel Text2Img Model Test with H100"
+  #   depends_on: upload-ready-pipeline
+  #   commands:
+  #     - |
+  #       timeout 30m bash -c '
+  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  #         pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py
+  #       '
+  #   agents:
+  #     queue: "mithril-h100-pool"
+  #   plugins:
+  #     - kubernetes:
+  #         podSpec:
+  #           containers:
+  #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+  #               resources:
+  #                 limits:
+  #                   nvidia.com/gpu: 1
+  #               volumeMounts:
+  #                 - name: devshm
+  #                   mountPath: /dev/shm
+  #                 - name: hf-cache
+  #                   mountPath: /root/.cache/huggingface
+  #               env:
+  #                 - name: HF_HOME
+  #                   value: /root/.cache/huggingface
+  #           nodeSelector:
+  #             node.kubernetes.io/instance-type: gpu-h100-sxm
+  #           volumes:
+  #             - name: devshm
+  #               emptyDir:
+  #                 medium: Memory
+  #             - name: hf-cache
+  #               hostPath:
+  #                 path: /mnt/hf-cache
+  #                 type: DirectoryOrCreate

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

@@ -14,7 +14,7 @@
         soft_fail: false
         commands:
           - "source .buildkite/scripts/docker_login_ecr_public.sh && safe_docker_login_ecr_public"
-          - "docker build -f docker/Dockerfile.rocm -t {{ docker_image_amd }} --target final --progress plain ."
+          - "docker build -f docker/Dockerfile.rocm -t {{ docker_image_amd }} --target test --build-arg ARG_PYTORCH_ROCM_ARCH=gfx942 --progress plain ."
           - "docker push {{ docker_image_amd }}"
         key: "amd-build"
         env:

.github/PULL_REQUEST_TEMPLATE.md

@@ -12,10 +12,10 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTT
 <summary> Essential Elements of an Effective PR Description Checklist </summary>
 
 - [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
-- [ ] The test plan, such as providing test command.
-- [ ] The test results, such as pasting the results comparison before and after, or e2e results
-- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model.
-- [ ] (Optional) Release notes update. If your change is user facing, please update the release notes draft.
+- [ ] The test plan. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the [test style doc](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/ci/tests_style/)
+- [ ] The test results. Please paste the results comparison before and after, or the e2e results.
+- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model. **Please run `mkdocs serve` to sync the documentation editions to `./docs`.**
+- [ ] (Optional) Release notes update. If your change is user-facing, please update the release notes draft.
 </details>
 
 **BEFORE SUBMITTING, PLEASE READ <https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md>** (anything written below this line will be removed by GitHub Actions)

.github/scripts/pr_reviewer.py

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

.github/workflows/comfyui-publish.yml

@@ -1,27 +0,0 @@
-name: ComfyUI Integration -- Publish to Comfy registry
-on:
-  workflow_dispatch:
-  push:
-    tags:
-      - "*"
-    paths:
-      - "apps/ComfyUI-vLLM-Omni/**"
-
-permissions:
-  issues: write
-
-defaults:
-  run:
-    working-directory: apps/ComfyUI-vLLM-Omni
-
-jobs:
-  publish-node:
-    name: Publish Custom Node to registry
-    runs-on: ubuntu-latest
-    steps:
-      - name: ♻️ Check out code
-        uses: actions/checkout@v4
-      - name: 📦 Publish Custom Node
-        uses: Comfy-Org/publish-node-action@main
-        with:
-          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

.github/workflows/comfyui-validate.yml

@@ -1,40 +0,0 @@
-name: ComfyUI Integration -- Validate backwards compatibility
-
-on:
-  pull_request:
-    branches:
-      - master
-      - main
-    paths:
-      - "apps/ComfyUI-vLLM-Omni/**"
-
-defaults:
-  run:
-    working-directory: apps/ComfyUI-vLLM-Omni
-
-jobs:
-  check-base-path:
-    runs-on: ubuntu-latest
-    outputs:
-      exists: ${{ steps.check.outputs.exists }}
-    defaults:
-      run:
-        working-directory: .
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event.pull_request.base.sha }}
-          fetch-depth: 1
-      - id: check
-        run: |
-          if [ -d "apps/ComfyUI-vLLM-Omni" ]; then
-            echo "exists=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "exists=false" >> "$GITHUB_OUTPUT"
-          fi
-  validate:
-    needs: check-base-path
-    if: needs.check-base-path.outputs.exists == 'true'
-    runs-on: ubuntu-latest
-    steps:
-      - uses: comfy-org/node-diff@main

.github/workflows/pr-reviewer.yml

@@ -0,0 +1,62 @@
+name: VLLM-Omni PR Reviewer
+
+on:
+  issue_comment:
+    types: [created]
+
+permissions:
+  contents: read
+  pull-requests: write
+  issues: write
+
+jobs:
+  pr-reviewer:
+    name: Review Pull Request
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    # Only run when the comment is from a collaborator/owner/member
+    if: |
+      github.event_name == 'issue_comment' &&
+      github.event.issue.pull_request != null &&
+      contains(github.event.comment.body, '@vllm-omni-reviewer') &&
+      (github.event.comment.author_association == 'MEMBER' ||
+       github.event.comment.author_association == 'COLLABORATOR' ||
+       github.event.comment.author_association == 'OWNER')
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4.2.2
+        with:
+          fetch-depth: 0
+
+      - name: Set up Python
+        uses: actions/setup-python@v5.3.0
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip==24.0
+          pip install requests==2.31.0 pyyaml==6.0.1
+
+      - name: Run PR Reviewer
+        id: reviewer
+        env:
+          GLM_API_KEY: ${{ secrets.GLM_API_KEY }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_NUMBER: ${{ github.event.issue.number || github.event.pull_request.number }}
+          COMMENT_BODY: ${{ github.event.comment.body }}
+          REPO_NAME: ${{ github.repository }}
+          PR_REVIEWER_COOLDOWN_MINUTES: 5
+          PR_REVIEWER_MAX_RETRIES: 3
+        run: |
+          python .github/scripts/pr_reviewer.py 2>&1 | tee "/tmp/pr_review_${PR_NUMBER}.log"
+
+      - name: Upload review logs
+        if: always()
+        uses: actions/upload-artifact@v4.4.3
+        with:
+          name: pr-review-logs-${{ github.event.issue.number || github.event.pull_request.number }}
+          path: /tmp/pr_review_*.log
+          retention-days: 7
+          if-no-files-found: ignore

.gitignore

@@ -242,3 +242,8 @@ configs/development.yaml
 Dockerfile.dev
 discussion
 tmp_test
+
+# Auto-generated version file (created by setuptools_scm during build)
+vllm_omni/_version.py
+# output files
+*.wav

apps/ComfyUI-vLLM-Omni/README.md

@@ -47,12 +47,12 @@ If no, check your shell running the ComfyUI process. There may be some error mes
 This extension offers the following nodes based on the output modalities (at **ComfyUI sidebar -> Node Library**):
 
 - **Generate Image** for text-to-image and image-to-image tasks
-- **Multimodality Comprehension** for multimodality-to-text and multimodality-to-audio tasks
+- **Multimodality Understanding** for multimodality-to-text and multimodality-to-audio tasks
 - **TTS** and **TTS Voice Clone** for TTS tasks
 
 This extension also offers example workflows (at **ComfyUI sidebar -> Templates -> vLLM-Omni**)
 
-> [!INFO]
+> [!NOTE]
 > The node UI and feature designs are intended to match vLLM-Omni online serving interfaces. It cannot offer more than what the interfaces support.
 
 To build a simple workflow yourself,
@@ -65,28 +65,16 @@ To build a simple workflow yourself,
     - For some multi-stage models like BAGEL, [only one stage's sampling parameters are exposed and tunable via vLLM-Omni's online serving API](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/examples/online_serving/bagel/). Thus, these models are treated as single-stage ones. Please check the vLLM-Omni documentation on how to correctly set each model's sampling parameters.
     - For multi-stage models where all stages are either autoregression or diffusion, you can also connect only a single Sampling Params node, indicating that this set of sampling parameters will be used for all stages.
 
-**The following features are tested**:
-
-- Single-node workflows for
-    - Multimodal Comprehension (e.g., Qwen Omni, BAGEL)
-    - Text-to-Image Generation (e.g., Qwen-Image)
-    - Image-to-Image Generation (e.g., Qwen-Image-Edit)
-    - TTS (e.g., Qwen TTS, including VoiceDesign, VoiceClone, CustomVoice)
-
-**The following features are not currently tested**. They may work or break. You are welcomed to test it out and offer comments.
-
-- Multi-node workflow that connects multiple model services together.
-
 ## Screenshots and Examples
 
-### Multimodal comprehension (e.g., Qwen Omni series, BAGEL)
+### Multimodal understanding (e.g., Qwen Omni series, BAGEL)
 
-(Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni Annotated Example**)
+(Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni Multimodal Understanding**)
 
 <p align="center">
   <picture>
-    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-comprehension.jpg">
-    <img alt="vLLM-Omni Main Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-comprehension.jpg" width=55%>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-understanding.jpg">
+    <img alt="vLLM-Omni multimodal understanding" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-understanding.jpg" width=55%>
   </picture>
 </p>
 
@@ -98,7 +86,7 @@ You can configure per-stage sampling parameters for multi-stage models.
 <p align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-multi-stage.jpg">
-    <img alt="vLLM-Omni Main Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-multi-stage.jpg" width=55%>
+    <img alt="vLLM-Omni multiple stages" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-multi-stage.jpg" width=55%>
   </picture>
 </p>
 
@@ -109,7 +97,7 @@ You can configure per-stage sampling parameters for multi-stage models.
 <p align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-image-generation.jpg">
-    <img alt="vLLM-Omni Main Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-image-generation.jpg" width=55%>
+    <img alt="vLLM-Omni image generation" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-image-generation.jpg" width=55%>
   </picture>
 </p>
 
@@ -123,13 +111,24 @@ You can configure per-stage sampling parameters for multi-stage models.
 <p align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-tts.jpg">
-    <img alt="vLLM-Omni Main Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-tts.jpg" width=55%>
+    <img alt="vLLM-Omni TTS" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-tts.jpg" width=55%>
   </picture>
 </p>
 
 > [!TIP]
 > There is a dedicated node for VoiceClone tasks with reference audio input. Other simple text-to-speech tasks should use the regular TTS node.
 
+### Chaining multiple model services
+
+(Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni Chaining Services**)
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-chaining-services.jpg">
+    <img alt="vLLM-Omni TTS" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-chaining-services.jpg" width=55%>
+  </picture>
+</p>
+
 ## Develop
 
 Follow the [development convention and rules of vLLM-Omni](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/).

apps/ComfyUI-vLLM-Omni/__init__.py

@@ -6,18 +6,18 @@ __all__ = [
     "WEB_DIRECTORY",
 ]
 
-__author__ = """Zeyu Huang"""
-__email__ = "11222265+fhfuih@users.noreply.github.com"
+__author__ = """vLLM-Omni Team"""
+__email__ = "vllm-omni@vllm.ai"
 __version__ = "0.0.1"
 
 from .comfyui_vllm_omni.nodes import (
     VLLMOmniARSampling,
-    VLLMOmniComprehension,
     VLLMOmniDiffusionSampling,
     VLLMOmniGenerateImage,
     VLLMOmniQwenTTSParams,
     VLLMOmniSamplingParamsList,
     VLLMOmniTTS,
+    VLLMOmniUnderstanding,
     VLLMOmniVoiceClone,
 )
 
@@ -25,7 +25,7 @@ from .comfyui_vllm_omni.nodes import (
 NODE_CLASS_MAPPINGS = {
     # === Generation ===
     "VLLMOmniGenerateImage": VLLMOmniGenerateImage,
-    "VLLMOmniComprehension": VLLMOmniComprehension,
+    "VLLMOmniUnderstanding": VLLMOmniUnderstanding,
     "VLLMOmniTTS": VLLMOmniTTS,
     "VLLMOmniVoiceClone": VLLMOmniVoiceClone,
     # === Params ===
@@ -39,7 +39,7 @@ NODE_CLASS_MAPPINGS = {
 NODE_DISPLAY_NAME_MAPPINGS = {
     # === Generation ===
     "VLLMOmniGenerateImage": "Generate Image",
-    "VLLMOmniComprehension": "Multimodality Comprehension",
+    "VLLMOmniUnderstanding": "Multimodality Understanding",
     "VLLMOmniTTS": "TTS (Text to Speech)",
     "VLLMOmniVoiceClone": "TTS Voice Cloning",
     # === Params ===

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/nodes.py

@@ -1,4 +1,4 @@
-from typing import Literal, cast
+from typing import Literal
 
 import torch
 from comfy_api.input import AudioInput, VideoInput
@@ -6,7 +6,12 @@ from comfy_api.input import AudioInput, VideoInput
 from .utils.api_client import VLLMOmniClient
 from .utils.logger import get_logger
 from .utils.models import lookup_model_spec
-from .utils.types import AudioFormat
+from .utils.types import (
+    AudioFormat,
+    AutoregressionSamplingParams,
+    DiffusionSamplingParams,
+    QwenTTSModelSpecificParams,
+)
 from .utils.validators import (
     add_sampling_parameters_to_stage,
     validate_model_and_sampling_params_types,
@@ -86,7 +91,10 @@ class VLLMOmniGenerateImage(_VLLMOmniGenerateBase):
 
         # Prefer DALL-E compatible API for simple (one-stage) diffusion models
         if (spec is None or spec["stages"] == ["diffusion"]) and not is_bagel:
-            sampling_params = cast(dict | None, sampling_params)
+            # The number of sampling parameter groups should have been validated.
+            # Now, simply convert single-item list to dict.
+            if isinstance(sampling_params, list):
+                sampling_params = sampling_params[0]
             if audio is None and image is None and video is None:
                 # No multimodal input --- use DALL-E image generation
                 logger.info("Using DALL-E image generation endpoint")
@@ -133,7 +141,7 @@ class VLLMOmniGenerateImage(_VLLMOmniGenerateBase):
         return (output,)
 
 
-class VLLMOmniComprehension(_VLLMOmniGenerateBase):
+class VLLMOmniUnderstanding(_VLLMOmniGenerateBase):
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -197,7 +205,7 @@ class VLLMOmniComprehension(_VLLMOmniGenerateBase):
             (
                 text_response,
                 _,
-            ) = await client.generate_comprehension_chat_completion(
+            ) = await client.generate_understanding_chat_completion(
                 model=model,
                 prompt=prompt,
                 image=image,
@@ -221,7 +229,7 @@ class VLLMOmniComprehension(_VLLMOmniGenerateBase):
             (
                 text_response,
                 audio,
-            ) = await client.generate_comprehension_chat_completion(
+            ) = await client.generate_understanding_chat_completion(
                 model=model,
                 prompt=prompt,
                 image=image,
@@ -287,15 +295,13 @@ class VLLMOmniTTS(_VLLMOmniGenerateBase):
         logger.info("Got extra kwargs in TTS: %s", kwargs)
 
         is_qwen_tts = "qwen3-tts" in model.lower()
-        extra_params_type = None if model_specific_params is None else model_specific_params["type"]
-        if not is_qwen_tts and extra_params_type == "qwen-tts":
+        if not is_qwen_tts and isinstance(model_specific_params, QwenTTSModelSpecificParams):
             raise ValueError(
                 "You have provided Qwen-specific TTS params."
                 "However, the model appears to not be a Qwen TTS model (no 'Qwen3-TTS' in model name)."
             )
 
         combined_params = {**kwargs, **(model_specific_params or {})}
-        combined_params.pop("type", None)  # Internal fields in model_specific_params
 
         client = VLLMOmniClient(url)
 
@@ -352,8 +358,7 @@ class VLLMOmniVoiceClone(_VLLMOmniGenerateBase):
         **kwargs,
     ):
         is_qwen_tts = "qwen3-tts" in model.lower()
-        extra_params_type = None if model_specific_params is None else model_specific_params["type"]
-        if not is_qwen_tts and extra_params_type == "qwen-tts":
+        if not is_qwen_tts and isinstance(model_specific_params, QwenTTSModelSpecificParams):
             raise ValueError(
                 "You have provided Qwen-specific TTS params."
                 "However, the model appears to not be a Qwen TTS model (no 'Qwen3-TTS' in model name)."
@@ -366,7 +371,6 @@ class VLLMOmniVoiceClone(_VLLMOmniGenerateBase):
             **kwargs,
             **(model_specific_params or {}),
         }
-        combined_params.pop("type", None)  # Internal fields in model_specific_params
 
         client = VLLMOmniClient(url)
 
@@ -419,10 +423,7 @@ class VLLMOmniARSampling:
     CATEGORY = "vLLM-Omni/Sampling Params"
 
     def get_params(self, seed, **kwargs):
-        params = {
-            "type": "autoregression",  # for internal use, removed before sending the request
-            **kwargs,
-        }
+        params = AutoregressionSamplingParams(kwargs)
         if seed >= 0:
             params["seed"] = seed
         return (params,)
@@ -479,6 +480,13 @@ class VLLMOmniDiffusionSampling:
                         "tooltip": "Enable VAE slicing for reduced memory usage (slight quality trade-off)",
                     },
                 ),
+                "vae_use_tiling": (
+                    "BOOLEAN",
+                    {
+                        "default": False,
+                        "tooltip": "Enable VAE tiling for reduced memory usage (slight quality trade-off)",
+                    },
+                ),
                 # === Put seed at last. ===
                 # Whenever a field named "seed" is present, ComfyUI adds another field called "control after generate"
                 "seed": (
@@ -499,10 +507,7 @@ class VLLMOmniDiffusionSampling:
     CATEGORY = "vLLM-Omni/Sampling Params"
 
     def get_params(self, seed, **kwargs):
-        params = {
-            "type": "diffusion",  # for internal use, removed before sending the request
-            **kwargs,
-        }
+        params = DiffusionSamplingParams(kwargs)
         if seed >= 0:
             params["seed"] = seed
         return (params,)
@@ -566,4 +571,4 @@ class VLLMOmniQwenTTSParams:
     CATEGORY = "vLLM-Omni/TTS Params"
 
     def get_params(self, **kwargs):
-        return ({"type": "qwen-tts", **kwargs},)
+        return (QwenTTSModelSpecificParams(kwargs),)

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/api_client.py

@@ -57,16 +57,7 @@ class VLLMOmniClient:
         if negative_prompt:
             payload["negative_prompt"] = negative_prompt
         if sampling_params is not None:
-            # Only select specific sampling params
-            for k in (
-                "n",
-                "num_inference_steps",
-                "guidance_scale",
-                "true_cfg_scale",
-                "vae_use_slicing",
-            ):
-                if k in sampling_params and sampling_params[k] is not None:
-                    payload[k] = sampling_params[k]
+            payload.update(sampling_params)
         logger.debug("img gen payload: %s", payload)
 
         url = self.base_url + "/images/generations"
@@ -138,10 +129,8 @@ class VLLMOmniClient:
         if negative_prompt:
             form.add_field("negative_prompt", negative_prompt)
         if sampling_params is not None:
-            # Only select specific sampling params
-            for k in ("n", "num_inference_steps", "guidance_scale", "true_cfg_scale"):
-                if k in sampling_params and sampling_params[k] is not None:
-                    form.add_field(k, str(sampling_params[k]))
+            for k, v in sampling_params.items():
+                form.add_field(k, str(v))
         if mask is not None:
             mask_filename = "mask.png"
             form.add_field(
@@ -217,7 +206,7 @@ class VLLMOmniClient:
 
         return torch.stack(image_tensors, dim=0)
 
-    async def generate_comprehension_chat_completion(
+    async def generate_understanding_chat_completion(
         self,
         *,
         model: str,
@@ -427,30 +416,28 @@ class VLLMOmniClient:
             message_content.append({"type": "video_url", "video_url": {"url": video_to_base64(video)}})
         messages = [{"role": "user", "content": message_content}]
 
+        payload: dict[str, Any] = {"messages": messages, "model": model}
+        if modalities:
+            payload["modalities"] = modalities
+
         combined_extra_body: dict[str, Any] = {}
         if sampling_params is not None:
             spec, _ = lookup_model_spec(model)
             is_single_sampling_param = isinstance(sampling_params, dict) or len(sampling_params) == 1
 
-            # Exclude internal key
-            if isinstance(sampling_params, dict):
-                sampling_params = {k: v for k, v in sampling_params.items() if k != "type"}
-            else:
-                sampling_params = [{k: v for k, v in sp.items() if k != "type"} for sp in sampling_params]
-
             if (spec is None and is_single_sampling_param) or (spec is not None and spec["stages"] == ["diffusion"]):
                 # Diffusion format: extra_body directly contains sampling params.
-                # Validation should have taken care of matching sampling params' types.
+                # Validation has already taken care of matching sampling params' types and length. Safe to take [0].
                 # * Use this mode if the model is a simple one-stage diffusion model.
                 # * Fallback to this mode if model is not registered and a single sampling param is provided.
                 sampling_params = sampling_params if isinstance(sampling_params, dict) else sampling_params[0]
-                combined_extra_body: dict[str, Any] = {**sampling_params}
+                combined_extra_body: dict[str, Any] = sampling_params.copy()
                 if "n" in combined_extra_body:
-                    combined_extra_body["num_outputs_per_prompt"] = combined_extra_body["n"]
-                    del combined_extra_body["n"]
+                    combined_extra_body["num_outputs_per_prompt"] = combined_extra_body.pop("n")
             else:
-                # Use AR style payload, extra_body has a sampling_params_list field
-                combined_extra_body: dict[str, Any] = {"sampling_params_list": sampling_params}
+                # AR format: the payload has a sampling_params_list field, containing a list.
+                sampling_params_list = sampling_params if isinstance(sampling_params, list) else [sampling_params]
+                payload["sampling_params_list"] = sampling_params_list
 
         if negative_prompt:
             combined_extra_body["negative_prompt"] = negative_prompt
@@ -458,12 +445,11 @@ class VLLMOmniClient:
         if extra_body:
             combined_extra_body.update(extra_body)
 
-        payload: dict[str, Any] = {"messages": messages, "model": model}
+        # Add extra_body only if it has any content.
         if combined_extra_body:
             payload["extra_body"] = combined_extra_body
-        if modalities:
-            payload["modalities"] = modalities
 
+        # Place to inject any model-specific payload adjustment
         spec, _ = lookup_model_spec(model)
         if spec:
             preprocessor = spec.get("payload_preprocessor", None)

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/models.py

@@ -11,8 +11,7 @@ def _bagel_payload_preprocessor(payload: dict) -> dict:
                     content["text"] = "<|im_start|>" + content["text"] + "<|im_end|>"
     except (KeyError, TypeError):
         raise RuntimeError("Internal Error: malformatted BAGEL payload")
-    extra_body = payload.pop("extra_body", {})
-    return {**payload, **extra_body}
+    return payload
 
 
 def _qwen25_payload_preprocessor(payload: dict) -> dict:
@@ -37,7 +36,7 @@ _MODEL_PIPELINE_SPECS: dict[str, Spec] = {
         ],
         "modes": [
             {
-                "mode": ModelMode.COMPREHENSION,
+                "mode": ModelMode.UNDERSTANDING,
                 "input_modalities": [Modality.TEXT, Modality.IMAGE],
             }
         ],
@@ -48,7 +47,7 @@ _MODEL_PIPELINE_SPECS: dict[str, Spec] = {
         "payload_preprocessor": _qwen25_payload_preprocessor,
         "modes": [
             {
-                "mode": ModelMode.COMPREHENSION,
+                "mode": ModelMode.UNDERSTANDING,
                 "input_modalities": [
                     Modality.TEXT,
                     Modality.IMAGE,
@@ -62,7 +61,7 @@ _MODEL_PIPELINE_SPECS: dict[str, Spec] = {
         "stages": ["autoregression", "autoregression", "autoregression"],
         "modes": [
             {
-                "mode": ModelMode.COMPREHENSION,
+                "mode": ModelMode.UNDERSTANDING,
                 "input_modalities": [
                     Modality.TEXT,
                     Modality.IMAGE,

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/types.py

@@ -11,11 +11,23 @@ from typing import (
 AudioFormat: TypeAlias = Literal["mp3", "opus", "aac", "flac", "wav", "pcm"]
 
 
+class AutoregressionSamplingParams(dict):
+    pass
+
+
+class DiffusionSamplingParams(dict):
+    pass
+
+
+class QwenTTSModelSpecificParams(dict):
+    pass
+
+
 class ModelMode(Enum):
     IMAGE_GENERATION = auto()
     VIDEO_GENERATION = auto()
     AUDIO_GENERATION = auto()
-    COMPREHENSION = auto()
+    UNDERSTANDING = auto()
 
 
 class Modality(Enum):

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/validators.py

@@ -1,5 +1,6 @@
 from .logger import get_logger
 from .models import lookup_model_spec
+from .types import AutoregressionSamplingParams, DiffusionSamplingParams
 
 logger = get_logger(__name__)
 
@@ -31,24 +32,20 @@ def validate_model_and_sampling_params_types(
             )
         # Check that each stage's type match
         for i, sp in enumerate(sampling_param_list):
-            if "type" not in sp:
-                raise RuntimeError("Internal error: unknown sampling parameter type")
-            if sp["type"] != stages[i]:
+            if not _check_sampling_param_matches_stage(sp, stages[i]):
                 raise ValueError(
-                    f"Sampling parameter type ({sp['type']}) does not match "
+                    f"Sampling parameter type ({sp.__class__.__name__}) does not match "
                     f"stage type ({stages[i]}) at index {i} for model {model_name}."
                 )
     elif isinstance(sampling_param_list, dict):
-        if "type" not in sampling_param_list:
-            raise RuntimeError("Internal error: unknown sampling parameter type")
         # Check that the provided single sampling param matches all stages
-        elif any(stage != sampling_param_list["type"] for stage in stages):
-            raise ValueError(
-                f"When passing a single sampling parameter node, all stages of the model must match "
-                f"the provided sampling parameter's type. "
-                f"However, the stages of model {model_name} are: {stages}. "
-                f"The provided sampling parameter is {sampling_param_list['type']}"
-            )
+        for i, stage in enumerate(stages):
+            if not _check_sampling_param_matches_stage(sampling_param_list, stage):
+                raise ValueError(
+                    f"Provided single sampling parameter type ({sampling_param_list.__class__.__name__}) must match "
+                    f"the types of all stages of the model. "
+                    f"However, stage {i} of model {model_name} is of type {stage}."
+                )
 
 
 def add_sampling_parameters_to_stage(
@@ -82,3 +79,11 @@ def add_sampling_parameters_to_stage(
                 sampling_param_list[i].update(params_to_add)
 
     return sampling_param_list
+
+
+def _check_sampling_param_matches_stage(sampling_param: dict, stage_type: str) -> bool:
+    if stage_type == "autoregression":
+        return isinstance(sampling_param, AutoregressionSamplingParams)
+    if stage_type == "diffusion":
+        return isinstance(sampling_param, DiffusionSamplingParams)
+    raise RuntimeError(f"Internal error: unknown stage type {stage_type}.")

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni Annotated Example.json

@@ -1 +0,0 @@
-{"id":"1c99f525-0a37-45ba-a28a-7df7c3af66b4","revision":0,"last_node_id":12,"last_link_id":14,"nodes":[{"id":1,"type":"VLLMOmniComprehension","pos":[1191.2177053682556,144.66829928181377],"size":[400,268],"flags":{},"order":8,"mode":0,"inputs":[{"localized_name":"image","name":"image","shape":7,"type":"IMAGE","link":1},{"localized_name":"video","name":"video","shape":7,"type":"VIDEO","link":2},{"localized_name":"audio","name":"audio","shape":7,"type":"AUDIO","link":3},{"localized_name":"sampling_params","name":"sampling_params","shape":7,"type":"SAMPLING_PARAMS","link":14},{"localized_name":"url","name":"url","type":"STRING","widget":{"name":"url"},"link":null},{"localized_name":"model","name":"model","type":"STRING","widget":{"name":"model"},"link":null},{"localized_name":"prompt","name":"prompt","type":"STRING","widget":{"name":"prompt"},"link":null},{"localized_name":"output_text","name":"output_text","type":"BOOLEAN","widget":{"name":"output_text"},"link":null},{"localized_name":"output_audio","name":"output_audio","type":"BOOLEAN","widget":{"name":"output_audio"},"link":null},{"localized_name":"use_audio_in_video","name":"use_audio_in_video","type":"BOOLEAN","widget":{"name":"use_audio_in_video"},"link":null}],"outputs":[{"localized_name":"text_response","name":"text_response","type":"STRING","links":[8]},{"localized_name":"audio_response","name":"audio_response","type":"AUDIO","links":[9]}],"properties":{"Node name for S&R":"VLLMOmniComprehension"},"widgets_values":["http://localhost:8000/v1","Qwen/Qwen2.5-Omni-7B","",true,true,true]},{"id":3,"type":"LoadVideo","pos":[729.5984141255855,-198.631920454299],"size":[282.798828125,233.0743408203125],"flags":{},"order":0,"mode":0,"inputs":[{"localized_name":"file","name":"file","type":"COMBO","widget":{"name":"file"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"VIDEO","name":"VIDEO","type":"VIDEO","links":[2]}],"properties":{"Node name for S&R":"LoadVideo"},"widgets_values":["draw.mp4","image"]},{"id":4,"type":"LoadAudio","pos":[729.8037086965753,99.86963519703949],"size":[282.798828125,136],"flags":{},"order":1,"mode":0,"inputs":[{"localized_name":"audio","name":"audio","type":"COMBO","widget":{"name":"audio"},"link":null},{"localized_name":"audioUI","name":"audioUI","type":"AUDIO_UI","widget":{"name":"audioUI"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"AUDIOUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"AUDIO","name":"AUDIO","type":"AUDIO","links":[3]}],"properties":{"Node name for S&R":"LoadAudio"},"widgets_values":["Megan-Fox.mp3",null,null]},{"id":5,"type":"VLLMOmniARSampling","pos":[510.3517536642828,658.073751009259],"size":[270,178],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"max_tokens","name":"max_tokens","type":"INT","widget":{"name":"max_tokens"},"link":null},{"localized_name":"temperature","name":"temperature","type":"FLOAT","widget":{"name":"temperature"},"link":null},{"localized_name":"top_p","name":"top_p","type":"FLOAT","widget":{"name":"top_p"},"link":null},{"localized_name":"repetition_penalty","name":"repetition_penalty","type":"FLOAT","widget":{"name":"repetition_penalty"},"link":null},{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":null}],"outputs":[{"localized_name":"AR sampling params","name":"AR sampling params","type":"SAMPLING_PARAMS","links":[12]}],"properties":{"Node name for S&R":"VLLMOmniARSampling"},"widgets_values":[100,1,1,1,-1,"randomize"]},{"id":7,"type":"VLLMOmniARSampling","pos":[503.33235181647115,419.34158016181806],"size":[270,178],"flags":{},"order":3,"mode":0,"inputs":[{"localized_name":"max_tokens","name":"max_tokens","type":"INT","widget":{"name":"max_tokens"},"link":null},{"localized_name":"temperature","name":"temperature","type":"FLOAT","widget":{"name":"temperature"},"link":null},{"localized_name":"top_p","name":"top_p","type":"FLOAT","widget":{"name":"top_p"},"link":null},{"localized_name":"repetition_penalty","name":"repetition_penalty","type":"FLOAT","widget":{"name":"repetition_penalty"},"link":null},{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":null}],"outputs":[{"localized_name":"AR sampling params","name":"AR sampling params","type":"SAMPLING_PARAMS","links":[5,13]}],"properties":{"Node name for S&R":"VLLMOmniARSampling"},"widgets_values":[100,1,1,1,-1,"randomize"]},{"id":8,"type":"VLLMOmniSamplingParamsList","pos":[820.6056617389042,426.38372037182273],"size":[263.066015625,66],"flags":{},"order":7,"mode":0,"inputs":[{"localized_name":"param1","name":"param1","type":"SAMPLING_PARAMS","link":5},{"localized_name":"param2","name":"param2","shape":7,"type":"SAMPLING_PARAMS","link":12},{"localized_name":"param3","name":"param3","shape":7,"type":"SAMPLING_PARAMS","link":13}],"outputs":[{"localized_name":"param list","name":"param list"Line truncated

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni Chaining Services.json

@@ -0,0 +1,552 @@
+{
+    "id": "6643e5fd-fa2a-4f25-935a-483173a8097c",
+    "revision": 0,
+    "last_node_id": 8,
+    "last_link_id": 6,
+    "nodes": [
+        {
+            "id": 1,
+            "type": "PreviewImage",
+            "pos": [
+                1446.2005205859512,
+                -316.7359686049902
+            ],
+            "size": [
+                305.6628685610174,
+                307.0336884172169
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "images",
+                    "name": "images",
+                    "type": "IMAGE",
+                    "link": 1
+                }
+            ],
+            "outputs": [],
+            "properties": {
+                "Node name for S&R": "PreviewImage"
+            },
+            "widgets_values": []
+        },
+        {
+            "id": 8,
+            "type": "PreviewImage",
+            "pos": [
+                1010.1848219273943,
+                -311.54216706114363
+            ],
+            "size": [
+                305.6628685610174,
+                307.0336884172169
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "images",
+                    "name": "images",
+                    "type": "IMAGE",
+                    "link": 6
+                }
+            ],
+            "outputs": [],
+            "properties": {
+                "Node name for S&R": "PreviewImage"
+            },
+            "widgets_values": []
+        },
+        {
+            "id": 7,
+            "type": "VLLMOmniDiffusionSampling",
+            "pos": [
+                234.61958293819296,
+                9.405681270043619
+            ],
+            "size": [
+                284.205078125,
+                226
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "n",
+                    "name": "n",
+                    "type": "INT",
+                    "widget": {
+                        "name": "n"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "num_inference_steps",
+                    "name": "num_inference_steps",
+                    "type": "INT",
+                    "widget": {
+                        "name": "num_inference_steps"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "guidance_scale",
+                    "name": "guidance_scale",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "guidance_scale"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "true_cfg_scale",
+                    "name": "true_cfg_scale",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "true_cfg_scale"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "vae_use_slicing",
+                    "name": "vae_use_slicing",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "vae_use_slicing"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "vae_use_tiling",
+                    "name": "vae_use_tiling",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "vae_use_tiling"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "seed",
+                    "name": "seed",
+                    "type": "INT",
+                    "widget": {
+                        "name": "seed"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "diffusion sampling params",
+                    "name": "diffusion sampling params",
+                    "type": "SAMPLING_PARAMS",
+                    "links": [
+                        5
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniDiffusionSampling"
+            },
+            "widgets_values": [
+                1,
+                50,
+                1,
+                1,
+                false,
+                false,
+                1525,
+                "randomize"
+            ]
+        },
+        {
+            "id": 4,
+            "type": "VLLMOmniDiffusionSampling",
+            "pos": [
+                666.8380026154548,
+                268.86271068330126
+            ],
+            "size": [
+                284.205078125,
+                226
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "n",
+                    "name": "n",
+                    "type": "INT",
+                    "widget": {
+                        "name": "n"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "num_inference_steps",
+                    "name": "num_inference_steps",
+                    "type": "INT",
+                    "widget": {
+                        "name": "num_inference_steps"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "guidance_scale",
+                    "name": "guidance_scale",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "guidance_scale"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "true_cfg_scale",
+                    "name": "true_cfg_scale",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "true_cfg_scale"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "vae_use_slicing",
+                    "name": "vae_use_slicing",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "vae_use_slicing"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "vae_use_tiling",
+                    "name": "vae_use_tiling",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "vae_use_tiling"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "seed",
+                    "name": "seed",
+                    "type": "INT",
+                    "widget": {
+                        "name": "seed"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "diffusion sampling params",
+                    "name": "diffusion sampling params",
+                    "type": "SAMPLING_PARAMS",
+                    "links": [
+                        2
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniDiffusionSampling"
+            },
+            "widgets_values": [
+                4,
+                50,
+                7,
+                1,
+                false,
+                false,
+                42,
+                "fixed"
+            ]
+        },
+        {
+            "id": 5,
+            "type": "VLLMOmniGenerateImage",
+            "pos": [
+                984.723613585788,
+                63.376900027553276
+            ],
+            "size": [
+                416.56628685610167,
+                372.1662621294205
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "image",
+                    "name": "image",
+                    "shape": 7,
+                    "type": "IMAGE",
+                    "link": 4
+                },
+                {
+                    "localized_name": "mask",
+                    "name": "mask",
+                    "shape": 7,
+                    "type": "MASK",
+                    "link": null
+                },
+                {
+                    "localized_name": "sampling_params",
+                    "name": "sampling_params",
+                    "shape": 7,
+                    "type": "SAMPLING_PARAMS",
+                    "link": 2
+                },
+                {
+                    "localized_name": "url",
+                    "name": "url",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "url"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "model",
+                    "name": "model",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "model"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "prompt",
+                    "name": "prompt",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "prompt"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "negative_prompt",
+                    "name": "negative_prompt",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "negative_prompt"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "width",
+                    "name": "width",
+                    "type": "INT",
+                    "widget": {
+                        "name": "width"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "height",
+                    "name": "height",
+                    "type": "INT",
+                    "widget": {
+                        "name": "height"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "image",
+                    "name": "image",
+                    "type": "IMAGE",
+                    "links": [
+                        1
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniGenerateImage"
+            },
+            "widgets_values": [
+                "http://localhost:8001/v1",
+                "/home/models/Qwen/Qwen-Image-Edit",
+                "A high-quality, high contrast, stylized portrait of the object in the uploaded reference image. Pop art and doodle style, with abundant scribbling patterns, such as a teal crown, orange lightning bolts, colorful handwritten scripts.",
+                "Realistic",
+                800,
+                800
+            ]
+        },
+        {
+            "id": 6,
+            "type": "VLLMOmniGenerateImage",
+            "pos": [
+                541.2702717429818,
+                -160.70392744708548
+            ],
+            "size": [
+                416.56628685610167,
+                372.1662621294205
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "image",
+                    "name": "image",
+                    "shape": 7,
+                    "type": "IMAGE",
+                    "link": null
+                },
+                {
+                    "localized_name": "mask",
+                    "name": "mask",
+                    "shape": 7,
+                    "type": "MASK",
+                    "link": null
+                },
+                {
+                    "localized_name": "sampling_params",
+                    "name": "sampling_params",
+                    "shape": 7,
+                    "type": "SAMPLING_PARAMS",
+                    "link": 5
+                },
+                {
+                    "localized_name": "url",
+                    "name": "url",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "url"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "model",
+                    "name": "model",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "model"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "prompt",
+                    "name": "prompt",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "prompt"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "negative_prompt",
+                    "name": "negative_prompt",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "negative_prompt"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "width",
+                    "name": "width",
+                    "type": "INT",
+                    "widget": {
+                        "name": "width"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "height",
+                    "name": "height",
+                    "type": "INT",
+                    "widget": {
+                        "name": "height"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "image",
+                    "name": "image",
+                    "type": "IMAGE",
+                    "links": [
+                        4,
+                        6
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniGenerateImage"
+            },
+            "widgets_values": [
+                "http://localhost:8000/v1",
+                "/home/models/Tongyi-MAI/Z-Image-Turbo",
+                "A headshot of a cute Siamese kitty. Blurred background due to wide aperture. Close-up look. Realistic.",
+                "Cartoonish.",
+                800,
+                800
+            ]
+        }
+    ],
+    "links": [
+        [
+            1,
+            5,
+            0,
+            1,
+            0,
+            "IMAGE"
+        ],
+        [
+            2,
+            4,
+            0,
+            5,
+            2,
+            "SAMPLING_PARAMS"
+        ],
+        [
+            4,
+            6,
+            0,
+            5,
+            0,
+            "IMAGE"
+        ],
+        [
+            5,
+            7,
+            0,
+            6,
+            2,
+            "SAMPLING_PARAMS"
+        ],
+        [
+            6,
+            6,
+            0,
+            8,
+            0,
+            "IMAGE"
+        ]
+    ],
+    "groups": [],
+    "config": {},
+    "extra": {
+        "workflowRendererVersion": "LG",
+        "ds": {
+            "scale": 1.0426432563169903,
+            "offset": [
+                -90.4071374799346,
+                575.2948920069742
+            ]
+        }
+    },
+    "version": 0.4
+}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni Image Generation.json

@@ -1 +1 @@
-{"id":"91f75acc-8040-40f6-865a-2e8a7cfd6672","revision":0,"last_node_id":11,"last_link_id":23,"nodes":[{"id":3,"type":"PreviewImage","pos":[1281.8455167767304,-69.02638461454333],"size":[305.6628685610174,307.0336884172169],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"images","name":"images","type":"IMAGE","link":22}],"outputs":[],"properties":{"Node name for S&R":"PreviewImage"},"widgets_values":[]},{"id":11,"type":"VLLMOmniGenerateImage","pos":[816.0962515873642,-137.51112584854445],"size":[400,278],"flags":{},"order":3,"mode":0,"inputs":[{"localized_name":"image","name":"image","shape":7,"type":"IMAGE","link":21},{"localized_name":"mask","name":"mask","shape":7,"type":"MASK","link":null},{"localized_name":"sampling_params","name":"sampling_params","shape":7,"type":"SAMPLING_PARAMS","link":23},{"localized_name":"url","name":"url","type":"STRING","widget":{"name":"url"},"link":null},{"localized_name":"model","name":"model","type":"STRING","widget":{"name":"model"},"link":null},{"localized_name":"prompt","name":"prompt","type":"STRING","widget":{"name":"prompt"},"link":null},{"localized_name":"negative_prompt","name":"negative_prompt","type":"STRING","widget":{"name":"negative_prompt"},"link":null},{"localized_name":"width","name":"width","type":"INT","widget":{"name":"width"},"link":null},{"localized_name":"height","name":"height","type":"INT","widget":{"name":"height"},"link":null}],"outputs":[{"localized_name":"image","name":"image","type":"IMAGE","links":[22]}],"properties":{"Node name for S&R":"VLLMOmniGenerateImage"},"widgets_values":["http://localhost:8000/v1","Qwen/Qwen-Image-Edit","Put this figure in a realistic mountain view","",512,512]},{"id":8,"type":"VLLMOmniDiffusionSampling","pos":[478.59266934006774,183.67711984955648],"size":[284.205078125,202],"flags":{},"order":0,"mode":0,"inputs":[{"localized_name":"n","name":"n","type":"INT","widget":{"name":"n"},"link":null},{"localized_name":"num_inference_steps","name":"num_inference_steps","type":"INT","widget":{"name":"num_inference_steps"},"link":null},{"localized_name":"guidance_scale","name":"guidance_scale","type":"FLOAT","widget":{"name":"guidance_scale"},"link":null},{"localized_name":"true_cfg_scale","name":"true_cfg_scale","type":"FLOAT","widget":{"name":"true_cfg_scale"},"link":null},{"localized_name":"vae_use_slicing","name":"vae_use_slicing","type":"BOOLEAN","widget":{"name":"vae_use_slicing"},"link":null},{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":null}],"outputs":[{"localized_name":"diffusion sampling params","name":"diffusion sampling params","type":"SAMPLING_PARAMS","links":[23]}],"properties":{"Node name for S&R":"VLLMOmniDiffusionSampling"},"widgets_values":[4,50,1,1,false,42,"fixed"]},{"id":10,"type":"MarkdownNote","pos":[227.99306819575278,-231.24143306069843],"size":[240.3326478922474,136.3505820791881],"flags":{},"order":2,"mode":0,"inputs":[],"outputs":[],"title":"Note: Task and Input","properties":{},"widgets_values":["vLLM-Omni nodes are categorized based on the output modality. The \"Generate Image\" node supports both text-to-image generation or image-to-image generation (a.k.a. image editing). The node will route to the correct endpoint depending on whether an input image is present or not."],"color":"#432","bgcolor":"#000"},{"id":4,"type":"LoadImage","pos":[496.31859627609606,-229.71277089860084],"size":[270,314],"flags":{},"order":1,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[21]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["example.png","image"]}],"links":[[21,4,0,11,0,"IMAGE"],[22,11,0,3,0,"IMAGE"],[23,8,0,11,2,"SAMPLING_PARAMS"]],"groups":[{"id":1,"title":"Input","bounding":[213.1706010087147,-313.20095750667554,560.8246471437027,407.4246532472182],"color":"#3f789e","font_size":24,"flags":{}}],"config":{},"extra":{"workflowRendererVersion":"LG","ds":{"scale":1.1469075819486894,"offset":[276.8427259980847,549.4885226647582]}},"version":0.4}
+{"id":"91f75acc-8040-40f6-865a-2e8a7cfd6672","revision":0,"last_node_id":11,"last_link_id":23,"nodes":[{"id":3,"type":"PreviewImage","pos":[1281.8455167767304,-69.02638461454333],"size":[305.6628685610174,307.0336884172169],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"图像","name":"images","type":"IMAGE","link":22}],"outputs":[],"properties":{"Node name for S&R":"PreviewImage"},"widgets_values":[]},{"id":11,"type":"VLLMOmniGenerateImage","pos":[816.0962515873642,-137.51112584854445],"size":[400,278],"flags":{},"order":3,"mode":0,"inputs":[{"localized_name":"image","name":"image","shape":7,"type":"IMAGE","link":21},{"localized_name":"mask","name":"mask","shape":7,"type":"MASK","link":null},{"localized_name":"sampling_params","name":"sampling_params","shape":7,"type":"SAMPLING_PARAMS","link":23},{"localized_name":"url","name":"url","type":"STRING","widget":{"name":"url"},"link":null},{"localized_name":"model","name":"model","type":"STRING","widget":{"name":"model"},"link":null},{"localized_name":"prompt","name":"prompt","type":"STRING","widget":{"name":"prompt"},"link":null},{"localized_name":"negative_prompt","name":"negative_prompt","type":"STRING","widget":{"name":"negative_prompt"},"link":null},{"localized_name":"width","name":"width","type":"INT","widget":{"name":"width"},"link":null},{"localized_name":"height","name":"height","type":"INT","widget":{"name":"height"},"link":null}],"outputs":[{"localized_name":"image","name":"image","type":"IMAGE","links":[22]}],"properties":{"Node name for S&R":"VLLMOmniGenerateImage"},"widgets_values":["http://localhost:8000/v1","Qwen/Qwen-Image-Edit","Put this figure in a realistic mountain view","",512,512]},{"id":10,"type":"MarkdownNote","pos":[227.99306819575278,-231.24143306069843],"size":[240.3326478922474,136.3505820791881],"flags":{},"order":1,"mode":0,"inputs":[],"outputs":[],"title":"Note: Task and Input","properties":{},"widgets_values":["vLLM-Omni nodes are categorized based on the output modality. The \"Generate Image\" node supports both text-to-image generation or image-to-image generation (a.k.a. image editing). The node will route to the correct endpoint depending on whether an input image is present or not."],"color":"#432","bgcolor":"#000"},{"id":4,"type":"LoadImage","pos":[496.31859627609606,-229.71277089860084],"size":[270,314],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"图像","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"选择文件上传","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"图像","name":"IMAGE","type":"IMAGE","links":[21]},{"localized_name":"遮罩","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["example.png","image"]},{"id":8,"type":"VLLMOmniDiffusionSampling","pos":[478.59266934006774,183.67711984955648],"size":[284.205078125,226],"flags":{},"order":0,"mode":0,"inputs":[{"localized_name":"n","name":"n","type":"INT","widget":{"name":"n"},"link":null},{"localized_name":"num_inference_steps","name":"num_inference_steps","type":"INT","widget":{"name":"num_inference_steps"},"link":null},{"localized_name":"guidance_scale","name":"guidance_scale","type":"FLOAT","widget":{"name":"guidance_scale"},"link":null},{"localized_name":"true_cfg_scale","name":"true_cfg_scale","type":"FLOAT","widget":{"name":"true_cfg_scale"},"link":null},{"localized_name":"vae_use_slicing","name":"vae_use_slicing","type":"BOOLEAN","widget":{"name":"vae_use_slicing"},"link":null},{"localized_name":"vae_use_tiling","name":"vae_use_tiling","type":"BOOLEAN","widget":{"name":"vae_use_tiling"},"link":null},{"localized_name":"seed","name":"seed","type":"INT","widget":{"name":"seed"},"link":null}],"outputs":[{"localized_name":"diffusion sampling params","name":"diffusion sampling params","type":"SAMPLING_PARAMS","links":[23]}],"properties":{"Node name for S&R":"VLLMOmniDiffusionSampling"},"widgets_values":[4,50,1,1,false,false,42,"randomize"]}],"links":[[21,4,0,11,0,"IMAGE"],[22,11,0,3,0,"IMAGE"],[23,8,0,11,2,"SAMPLING_PARAMS"]],"groups":[{"id":1,"title":"Input","bounding":[213.1706010087147,-313.20095750667554,560.8246471437027,407.4246532472182],"color":"#3f789e","font_size":24,"flags":{}}],"config":{},"extra":{"workflowRendererVersion":"LG","ds":{"scale":1.1469075819486894,"offset":[-55.18047513099182,220.2553505195962]}},"version":0.4}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni Multimodal Understanding.json

@@ -0,0 +1,761 @@
+{
+    "id": "1c99f525-0a37-45ba-a28a-7df7c3af66b4",
+    "revision": 0,
+    "last_node_id": 12,
+    "last_link_id": 14,
+    "nodes": [
+        {
+            "id": 1,
+            "type": "VLLMOmniUnderstanding",
+            "pos": [
+                1191.2177053682556,
+                144.66829928181377
+            ],
+            "size": [
+                400,
+                268
+            ],
+            "flags": {},
+            "order": 8,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "image",
+                    "name": "image",
+                    "shape": 7,
+                    "type": "IMAGE",
+                    "link": 1
+                },
+                {
+                    "localized_name": "video",
+                    "name": "video",
+                    "shape": 7,
+                    "type": "VIDEO",
+                    "link": 2
+                },
+                {
+                    "localized_name": "audio",
+                    "name": "audio",
+                    "shape": 7,
+                    "type": "AUDIO",
+                    "link": 3
+                },
+                {
+                    "localized_name": "sampling_params",
+                    "name": "sampling_params",
+                    "shape": 7,
+                    "type": "SAMPLING_PARAMS",
+                    "link": 14
+                },
+                {
+                    "localized_name": "url",
+                    "name": "url",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "url"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "model",
+                    "name": "model",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "model"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "prompt",
+                    "name": "prompt",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "prompt"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "output_text",
+                    "name": "output_text",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "output_text"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "output_audio",
+                    "name": "output_audio",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "output_audio"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "use_audio_in_video",
+                    "name": "use_audio_in_video",
+                    "type": "BOOLEAN",
+                    "widget": {
+                        "name": "use_audio_in_video"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "text_response",
+                    "name": "text_response",
+                    "type": "STRING",
+                    "links": [
+                        8
+                    ]
+                },
+                {
+                    "localized_name": "audio_response",
+                    "name": "audio_response",
+                    "type": "AUDIO",
+                    "links": [
+                        9
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniUnderstanding"
+            },
+            "widgets_values": [
+                "http://localhost:8000/v1",
+                "Qwen/Qwen2.5-Omni-7B",
+                "",
+                true,
+                true,
+                true
+            ]
+        },
+        {
+            "id": 3,
+            "type": "LoadVideo",
+            "pos": [
+                729.5984141255855,
+                -198.631920454299
+            ],
+            "size": [
+                282.798828125,
+                233.0743408203125
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "file",
+                    "name": "file",
+                    "type": "COMBO",
+                    "widget": {
+                        "name": "file"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "choose file to upload",
+                    "name": "upload",
+                    "type": "IMAGEUPLOAD",
+                    "widget": {
+                        "name": "upload"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "VIDEO",
+                    "name": "VIDEO",
+                    "type": "VIDEO",
+                    "links": [
+                        2
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "LoadVideo"
+            },
+            "widgets_values": [
+                "draw.mp4",
+                "image"
+            ]
+        },
+        {
+            "id": 4,
+            "type": "LoadAudio",
+            "pos": [
+                729.8037086965753,
+                99.86963519703949
+            ],
+            "size": [
+                282.798828125,
+                136
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "audio",
+                    "name": "audio",
+                    "type": "COMBO",
+                    "widget": {
+                        "name": "audio"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "audioUI",
+                    "name": "audioUI",
+                    "type": "AUDIO_UI",
+                    "widget": {
+                        "name": "audioUI"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "choose file to upload",
+                    "name": "upload",
+                    "type": "AUDIOUPLOAD",
+                    "widget": {
+                        "name": "upload"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "AUDIO",
+                    "name": "AUDIO",
+                    "type": "AUDIO",
+                    "links": [
+                        3
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "LoadAudio"
+            },
+            "widgets_values": [
+                "Megan-Fox.mp3",
+                null,
+                null
+            ]
+        },
+        {
+            "id": 5,
+            "type": "VLLMOmniARSampling",
+            "pos": [
+                510.3517536642828,
+                658.073751009259
+            ],
+            "size": [
+                270,
+                178
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "max_tokens",
+                    "name": "max_tokens",
+                    "type": "INT",
+                    "widget": {
+                        "name": "max_tokens"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "temperature",
+                    "name": "temperature",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "temperature"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "top_p",
+                    "name": "top_p",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "top_p"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "repetition_penalty",
+                    "name": "repetition_penalty",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "repetition_penalty"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "seed",
+                    "name": "seed",
+                    "type": "INT",
+                    "widget": {
+                        "name": "seed"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "AR sampling params",
+                    "name": "AR sampling params",
+                    "type": "SAMPLING_PARAMS",
+                    "links": [
+                        12
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniARSampling"
+            },
+            "widgets_values": [
+                100,
+                1,
+                1,
+                1,
+                -1,
+                "randomize"
+            ]
+        },
+        {
+            "id": 7,
+            "type": "VLLMOmniARSampling",
+            "pos": [
+                503.33235181647115,
+                419.34158016181806
+            ],
+            "size": [
+                270,
+                178
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "max_tokens",
+                    "name": "max_tokens",
+                    "type": "INT",
+                    "widget": {
+                        "name": "max_tokens"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "temperature",
+                    "name": "temperature",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "temperature"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "top_p",
+                    "name": "top_p",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "top_p"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "repetition_penalty",
+                    "name": "repetition_penalty",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "repetition_penalty"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "seed",
+                    "name": "seed",
+                    "type": "INT",
+                    "widget": {
+                        "name": "seed"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "AR sampling params",
+                    "name": "AR sampling params",
+                    "type": "SAMPLING_PARAMS",
+                    "links": [
+                        5,
+                        13
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniARSampling"
+            },
+            "widgets_values": [
+                100,
+                1,
+                1,
+                1,
+                -1,
+                "randomize"
+            ]
+        },
+        {
+            "id": 8,
+            "type": "VLLMOmniSamplingParamsList",
+            "pos": [
+                820.6056617389042,
+                426.38372037182273
+            ],
+            "size": [
+                263.066015625,
+                66
+            ],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "param1",
+                    "name": "param1",
+                    "type": "SAMPLING_PARAMS",
+                    "link": 5
+                },
+                {
+                    "localized_name": "param2",
+                    "name": "param2",
+                    "shape": 7,
+                    "type": "SAMPLING_PARAMS",
+                    "link": 12
+                },
+                {
+                    "localized_name": "param3",
+                    "name": "param3",
+                    "shape": 7,
+                    "type": "SAMPLING_PARAMS",
+                    "link": 13
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "param list",
+                    "name": "param list",
+                    "type": "SAMPLING_PARAMS",
+                    "links": [
+                        14
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "VLLMOmniSamplingParamsList"
+            },
+            "widgets_values": []
+        },
+        {
+            "id": 11,
+            "type": "MarkdownNote",
+            "pos": [
+                826.2328280438272,
+                569.1890318701705
+            ],
+            "size": [
+                333.8220435590464,
+                261.63596728060656
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [],
+            "title": "Note: Sampling Parameters",
+            "properties": {},
+            "widgets_values": [
+                "## Sampling Parameter Types\n\nThere are two types of sampling parameters: one for autoregression and one for diffusion.\nYou should ensure that you have chosen the correct type of sampling parameters for the model you request.\n\n## Stages & Shorthand\n\nFor multi-stage models such as Qwen Omni, you can either\n- connect one sampling parameter node, which is applied to all stages.\n- connect exactly the same number of sampling parameter nodes to a \"Multi-Stage Sampling Parameter List\" node, then connect this node to the primary request node.\n\nNote that this shorthand is intended to stay consistent with the [online serving API](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/examples/online_serving/qwen2_5_omni/)"
+            ],
+            "color": "#432",
+            "bgcolor": "#000"
+        },
+        {
+            "id": 12,
+            "type": "MarkdownNote",
+            "pos": [
+                378.9866207003777,
+                152.59550252215752
+            ],
+            "size": [
+                319.7287574247016,
+                107.15904081906785
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [],
+            "title": "Note: Input",
+            "properties": {},
+            "widgets_values": [
+                "Note that not all models support every modality as input. For example, `ByteDance-Seed/BAGEL-7B-MoT` in Multimodality Understanding mode only support text and image input.\n\nYou should ensure that the input are supported by the model. You can check the corresponding [online serving documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/examples/online_serving/bagel/) for confirmation."
+            ],
+            "color": "#432",
+            "bgcolor": "#000"
+        },
+        {
+            "id": 2,
+            "type": "LoadImage",
+            "pos": [
+                394.4674804308822,
+                -207.6987397548834
+            ],
+            "size": [
+                282.798828125,
+                314
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "image",
+                    "name": "image",
+                    "type": "COMBO",
+                    "widget": {
+                        "name": "image"
+                    },
+                    "link": null
+                },
+                {
+                    "localized_name": "choose file to upload",
+                    "name": "upload",
+                    "type": "IMAGEUPLOAD",
+                    "widget": {
+                        "name": "upload"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "IMAGE",
+                    "name": "IMAGE",
+                    "type": "IMAGE",
+                    "links": [
+                        1
+                    ]
+                },
+                {
+                    "localized_name": "MASK",
+                    "name": "MASK",
+                    "type": "MASK",
+                    "links": null
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "LoadImage"
+            },
+            "widgets_values": [
+                "example.png",
+                "image"
+            ]
+        },
+        {
+            "id": 10,
+            "type": "PreviewAudio",
+            "pos": [
+                1664.548345556043,
+                297.5921292054054
+            ],
+            "size": [
+                270,
+                88
+            ],
+            "flags": {},
+            "order": 10,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "audio",
+                    "name": "audio",
+                    "type": "AUDIO",
+                    "link": 9
+                },
+                {
+                    "localized_name": "audioUI",
+                    "name": "audioUI",
+                    "type": "AUDIO_UI",
+                    "widget": {
+                        "name": "audioUI"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [],
+            "properties": {
+                "Node name for S&R": "PreviewAudio"
+            },
+            "widgets_values": []
+        },
+        {
+            "id": 9,
+            "type": "ShowText|pysssss",
+            "pos": [
+                1649.2506875091847,
+                66.22823888292349
+            ],
+            "size": [
+                318.7188464232943,
+                173.38502269972975
+            ],
+            "flags": {},
+            "order": 9,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "text",
+                    "name": "text",
+                    "type": "STRING",
+                    "link": 8
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "STRING",
+                    "name": "STRING",
+                    "shape": 6,
+                    "type": "STRING",
+                    "links": null
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "ShowText|pysssss"
+            },
+            "widgets_values": []
+        }
+    ],
+    "links": [
+        [
+            1,
+            2,
+            0,
+            1,
+            0,
+            "IMAGE"
+        ],
+        [
+            2,
+            3,
+            0,
+            1,
+            1,
+            "VIDEO"
+        ],
+        [
+            3,
+            4,
+            0,
+            1,
+            2,
+            "AUDIO"
+        ],
+        [
+            5,
+            7,
+            0,
+            8,
+            0,
+            "SAMPLING_PARAMS"
+        ],
+        [
+            8,
+            1,
+            0,
+            9,
+            0,
+            "STRING"
+        ],
+        [
+            9,
+            1,
+            1,
+            10,
+            0,
+            "AUDIO"
+        ],
+        [
+            12,
+            5,
+            0,
+            8,
+            1,
+            "SAMPLING_PARAMS"
+        ],
+        [
+            13,
+            7,
+            0,
+            8,
+            2,
+            "SAMPLING_PARAMS"
+        ],
+        [
+            14,
+            8,
+            0,
+            1,
+            3,
+            "SAMPLING_PARAMS"
+        ]
+    ],
+    "groups": [
+        {
+            "id": 1,
+            "title": "Sampling Parameters",
+            "bounding": [
+                480.1649301181556,
+                341.08402513937995,
+                692.4113568972277,
+                510.48648853403665
+            ],
+            "color": "#3f789e",
+            "font_size": 24,
+            "flags": {}
+        },
+        {
+            "id": 2,
+            "title": "Input",
+            "bounding": [
+                344.2364817813528,
+                -287.5850484183313,
+                704.8832238768634,
+                559.5124009832894
+            ],
+            "color": "#3f789e",
+            "font_size": 24,
+            "flags": {}
+        }
+    ],
+    "config": {},
+    "extra": {
+        "workflowRendererVersion": "LG",
+        "ds": {
+            "scale": 0.9478575057427204,
+            "offset": [
+                33.81037136029557,
+                307.1974296197726
+            ]
+        }
+    },
+    "version": 0.4
+}

benchmarks/build_dataset/download_process_data_seedtts.md

@@ -56,7 +56,7 @@ python extract_tts_prompts.py -i meta.lst -o top100.txt -n 100
 **Options:**
 - `-i, --input`: Input metadata file (default: `meta.lst`)
 - `-o, --output`: Output prompts file (default: `prompts.txt`)
-- `-n, --num_lines`: Number of prompts to extract (required)
+- `-n, --num-lines`: Number of prompts to extract (required)
 
 ### 7. Clean Up (Optional)
 

benchmarks/build_dataset/extract_tts_prompts.py

@@ -56,7 +56,7 @@ def main():
         "-o", "--output", type=str, default="prompts.txt", help="Output txt file path (default: prompts.txt)"
     )
     parser.add_argument(
-        "-n", "--num_lines", type=int, required=True, help="Number of lines to extract from the beginning"
+        "-n", "--num-lines", type=int, required=True, help="Number of lines to extract from the beginning"
     )
 
     args = parser.parse_args()

benchmarks/diffusion/backends.py

@@ -0,0 +1,198 @@
+import base64
+import mimetypes
+import os
+import time
+import uuid
+from dataclasses import dataclass, field
+from typing import Any
+
+import aiohttp
+from tqdm import tqdm
+
+
+@dataclass
+class RequestFuncInput:
+    prompt: str
+    api_url: str
+    model: str
+    width: int | None = None
+    height: int | None = None
+    num_frames: int | None = None
+    num_inference_steps: int | None = None
+    seed: int | None = None
+    fps: int | None = None
+    timestamp: float | None = None
+    slo_ms: float | None = None
+    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()))
+
+
+@dataclass
+class RequestFuncOutput:
+    success: bool = False
+    latency: float = 0.0
+    error: str = ""
+    start_time: float = 0.0
+    response_body: dict[str, Any] = field(default_factory=dict)
+    peak_memory_mb: float = 0.0
+    slo_achieved: bool | None = None
+
+
+def _guess_mime_type(path: str) -> str:
+    mime, _ = mimetypes.guess_type(path)
+    return mime or "application/octet-stream"
+
+
+def _encode_image_as_data_url(path: str) -> str:
+    with open(path, "rb") as f:
+        encoded = base64.b64encode(f.read()).decode("utf-8")
+    mime = _guess_mime_type(path)
+    return f"data:{mime};base64,{encoded}"
+
+
+async def async_request_chat_completions(
+    input: RequestFuncInput,
+    session: aiohttp.ClientSession,
+    pbar: tqdm | None = None,
+) -> RequestFuncOutput:
+    output = RequestFuncOutput()
+    output.start_time = time.perf_counter()
+
+    extra_body = dict(input.extra_body)
+    if input.width and input.height:
+        extra_body.setdefault("height", input.height)
+        extra_body.setdefault("width", input.width)
+    if input.num_frames:
+        extra_body.setdefault("num_frames", input.num_frames)
+    if input.num_inference_steps:
+        extra_body.setdefault("num_inference_steps", input.num_inference_steps)
+    if input.seed is not None:
+        extra_body.setdefault("seed", input.seed)
+    if input.fps:
+        extra_body.setdefault("fps", input.fps)
+
+    if input.image_paths and len(input.image_paths) > 0:
+        content = []
+        if input.prompt:
+            content.append({"type": "text", "text": input.prompt})
+        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
+            content.append(
+                {
+                    "type": "image_url",
+                    "image_url": {"url": _encode_image_as_data_url(img_path)},
+                }
+            )
+        messages = [{"role": "user", "content": content}]
+    else:
+        messages = [{"role": "user", "content": input.prompt}]
+
+    payload = {
+        "model": input.model,
+        "messages": messages,
+    }
+    if extra_body:
+        payload["extra_body"] = extra_body
+
+    try:
+        async with session.post(input.api_url, json=payload) as response:
+            if response.status == 200:
+                resp_json = await response.json()
+                output.response_body = resp_json
+                output.success = True
+                if "peak_memory_mb" in resp_json:
+                    output.peak_memory_mb = resp_json["peak_memory_mb"]
+            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_openai_images(
+    input: RequestFuncInput,
+    session: aiohttp.ClientSession,
+    pbar: tqdm | None = None,
+) -> RequestFuncOutput:
+    """
+    Send request to OpenAI's /v1/images/generations endpoint.
+    """
+    output = RequestFuncOutput()
+    output.start_time = time.perf_counter()
+
+    # Build size string from width/height
+    width = input.width or 1024
+    height = input.height or 1024
+    size = f"{width}x{height}"
+
+    payload: dict[str, Any] = {
+        "model": input.model,
+        "prompt": input.prompt,
+        "n": 1,
+        "size": size,
+        "response_format": "b64_json",
+    }
+
+    # Add optional parameters
+    if input.seed is not None:
+        payload["seed"] = input.seed
+    if input.num_inference_steps is not None:
+        payload["num_inference_steps"] = input.num_inference_steps
+
+    # Add any extra body parameters
+    if input.extra_body:
+        for key, value in input.extra_body.items():
+            if key not in payload:
+                payload[key] = value
+
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": "Bearer EMPTY",
+    }
+
+    try:
+        async with session.post(input.api_url, json=payload, headers=headers) as response:
+            if response.status == 200:
+                resp_json = await response.json()
+                output.response_body = resp_json
+                output.success = True
+                # Check for usage/memory info if available
+                if "usage" in resp_json and "peak_memory_mb" in resp_json.get("usage", {}):
+                    output.peak_memory_mb = resp_json["usage"]["peak_memory_mb"]
+            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
+
+
+backends_function_mapping = {
+    "vllm-omni": (async_request_chat_completions, "/v1/chat/completions"),
+    "openai": (async_request_openai_images, "/v1/images/generations"),
+}

benchmarks/diffusion/diffusion_benchmark_serving.py

@@ -6,80 +6,60 @@
 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
+
 Usage:
-    # Video
+    # Video (vllm-omni backend)
     t2v:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --dataset vbench --task t2v --num-prompts 10 \
+        --backend vllm-omni --dataset vbench --task t2v --num-prompts 10 \
         --height 480 --width 640 --fps 16 --num-frames 80
 
     i2v:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --dataset vbench --task i2v --num-prompts 10
+        --backend vllm-omni --dataset vbench --task i2v --num-prompts 10
 
 
-    # Image
+    # Image (vllm-omni backend)
     t2i:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --dataset vbench --task t2i --num-prompts 10 \
+        --backend vllm-omni --dataset vbench --task t2i --num-prompts 10 \
         --height 1024 --width 1024
 
     i2i:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --dataset vbench --task i2i --num-prompts 10
+        --backend vllm-omni --dataset vbench --task i2i --num-prompts 10
+
+    # Image (openai backend)
+    t2i:
+    python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
+        --backend openai --dataset vbench --task t2i --num-prompts 10 \
+        --height 1024 --width 1024 --port 3000
 
 """
 
 import argparse
 import ast
 import asyncio
-import base64
 import glob
 import json
-import mimetypes
 import os
 import time
 import uuid
 from abc import ABC, abstractmethod
 from collections.abc import AsyncGenerator
-from dataclasses import dataclass, field, replace
+from dataclasses import replace
 from typing import Any
 
 import aiohttp
 import numpy as np
 import requests
+from backends import RequestFuncInput, RequestFuncOutput, backends_function_mapping
 from tqdm.asyncio import tqdm
 
 
-@dataclass
-class RequestFuncInput:
-    prompt: str
-    api_url: str
-    model: str
-    width: int | None = None
-    height: int | None = None
-    num_frames: int | None = None
-    num_inference_steps: int | None = None
-    seed: int | None = None
-    fps: int | None = None
-    timestamp: float | None = None
-    slo_ms: float | None = None
-    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()))
-
-
-@dataclass
-class RequestFuncOutput:
-    success: bool = False
-    latency: float = 0.0
-    error: str = ""
-    start_time: float = 0.0
-    response_body: dict[str, Any] = field(default_factory=dict)
-    peak_memory_mb: float = 0.0
-    slo_achieved: bool | None = None
-
-
 class BaseDataset(ABC):
     def __init__(self, args, api_url: str, model: str):
         self.args = args
@@ -684,92 +664,6 @@ async def iter_requests(
         yield req
 
 
-def _guess_mime_type(path: str) -> str:
-    mime, _ = mimetypes.guess_type(path)
-    return mime or "application/octet-stream"
-
-
-def _encode_image_as_data_url(path: str) -> str:
-    with open(path, "rb") as f:
-        encoded = base64.b64encode(f.read()).decode("utf-8")
-    mime = _guess_mime_type(path)
-    return f"data:{mime};base64,{encoded}"
-
-
-async def async_request_chat_completions(
-    input: RequestFuncInput,
-    session: aiohttp.ClientSession,
-    pbar: tqdm | None = None,
-) -> RequestFuncOutput:
-    output = RequestFuncOutput()
-    output.start_time = time.perf_counter()
-
-    extra_body = dict(input.extra_body)
-    if input.width and input.height:
-        extra_body.setdefault("height", input.height)
-        extra_body.setdefault("width", input.width)
-    if input.num_frames:
-        extra_body.setdefault("num_frames", input.num_frames)
-    if input.num_inference_steps:
-        extra_body.setdefault("num_inference_steps", input.num_inference_steps)
-    if input.seed is not None:
-        extra_body.setdefault("seed", input.seed)
-    if input.fps:
-        extra_body.setdefault("fps", input.fps)
-
-    if input.image_paths and len(input.image_paths) > 0:
-        content = []
-        if input.prompt:
-            content.append({"type": "text", "text": input.prompt})
-        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
-            content.append(
-                {
-                    "type": "image_url",
-                    "image_url": {"url": _encode_image_as_data_url(img_path)},
-                }
-            )
-        messages = [{"role": "user", "content": content}]
-    else:
-        messages = [{"role": "user", "content": input.prompt}]
-
-    payload = {
-        "model": input.model,
-        "messages": messages,
-    }
-    if extra_body:
-        payload["extra_body"] = extra_body
-
-    try:
-        async with session.post(input.api_url, json=payload) as response:
-            if response.status == 200:
-                resp_json = await response.json()
-                output.response_body = resp_json
-                output.success = True
-                if "peak_memory_mb" in resp_json:
-                    output.peak_memory_mb = resp_json["peak_memory_mb"]
-            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
-
-
 def calculate_metrics(
     outputs: list[RequestFuncOutput],
     total_duration: float,
@@ -848,9 +742,9 @@ async def benchmark(args):
     if args.base_url is None:
         args.base_url = f"http://{args.host}:{args.port}"
 
-    # Setup dataset (vLLM-Omni supports diffusion via /v1/chat/completions)
-    api_url = f"{args.base_url}/v1/chat/completions"
-    request_func = async_request_chat_completions
+    # Setup API URL and request function based on backend
+    request_func, api_url = backends_function_mapping[args.backend]
+    api_url = f"{args.base_url}{api_url}"
 
     if args.dataset == "vbench":
         dataset = VBenchDataset(args, api_url, args.model)
@@ -920,9 +814,16 @@ async def benchmark(args):
     # Calculate metrics
     metrics = calculate_metrics(outputs, total_duration, requests_list, args, args.slo)
 
+    # Add configuration info to metrics for JSON output
+    metrics["backend"] = args.backend
+    metrics["model"] = args.model
+    metrics["dataset"] = args.dataset
+    metrics["task"] = args.task
+
     print("\n{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=60, c="="))
 
     # Section 1: Configuration
+    print("{:<40} {:<15}".format("Backend:", args.backend))
     print("{:<40} {:<15}".format("Model:", args.model))
     print("{:<40} {:<15}".format("Dataset:", args.dataset))
     print("{:<40} {:<15}".format("Task:", args.task))
@@ -978,6 +879,13 @@ 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(
+        "--backend",
+        type=str,
+        default="vllm-omni",
+        choices=["vllm-omni", "openai"],
+        help="Backend to target the benchmark to.",
+    )
     parser.add_argument(
         "--dataset",
         type=str,

benchmarks/distributed/omni_connectors/README.md

@@ -0,0 +1,397 @@
+# RDMA Test Configuration Guide
+
+This document explains how to configure the RDMA environment and run tests for `MooncakeTransferEngineConnector`.
+
+## Table of Contents
+
+- [Docker Container Permissions](#docker-container-permissions)
+- [Single-Node Testing](#single-node-testing)
+- [Multi-Node Testing](#multi-node-testing)
+- [Running Tests](#running-tests)
+- [Cross-Node Testing](#cross-node-testing)
+- [Troubleshooting](#troubleshooting)
+
+---
+
+## Docker Container Permissions
+
+RDMA tests require access to InfiniBand/RoCE devices and system topology. Add the following permissions when running `docker run`.
+
+### Option 1: Minimal Permissions (Recommended)
+
+```bash
+docker run -it \
+    --cap-add=SYS_PTRACE \
+    --cap-add=IPC_LOCK \
+    --security-opt seccomp=unconfined \
+    --network=host \
+    --device=/dev/infiniband \
+    -v /sys/class/infiniband:/sys/class/infiniband:ro \
+    your-image:tag
+```
+
+Parameter explanation:
+- `--cap-add=SYS_PTRACE`: Allow reading system topology information
+- `--cap-add=IPC_LOCK`: Allow memory locking (required for RDMA memory registration)
+- `--security-opt seccomp=unconfined`: Disable seccomp restrictions
+- `--network=host`: Use host network (required for RDMA)
+- `--device=/dev/infiniband`: Mount InfiniBand devices
+- `-v /sys/class/infiniband`: Mount IB device info (read-only)
+
+### Option 2: Full Permissions (Quick but not recommended for production)
+
+```bash
+docker run -it \
+    --privileged \
+    --network=host \
+    your-image:tag
+```
+
+`--privileged` grants full host permissions. Suitable for quick testing but not recommended for production.
+
+---
+
+## Single-Node Testing
+
+When running single-node tests (producer and consumer on the same machine), ensure they use the **same RDMA device**.
+
+### Problem Background
+
+InfiniBand devices use LID (Local Identifier) for routing. Different devices have different LIDs and cannot communicate directly. If no device is specified, Mooncake may assign different devices to connectors, causing handshake failures.
+
+Common error:
+```
+[Handshake] Failed to modify QP to RTR, check mtu, gid, peer lid, peer qp num: Invalid argument [22]
+```
+
+### Solution
+
+**Method 1: Set Environment Variable (Recommended)**
+
+```bash
+# List available RDMA devices
+ibstat
+
+# Select a device (e.g., mlx5_0)
+export RDMA_DEVICE_NAME='mlx5_0'
+
+# Run tests
+pytest test_mooncake_transfer_engine_rdma.py -v -s
+```
+
+**Method 2: Use RoCE Devices**
+
+If the system has RoCE devices (using IPv4 routing), the test code will automatically detect and prefer them. RoCE device GIDs start with `00:00:00:00:00:00:00:00:00:00:ff:ff` (IPv4-mapped).
+
+**Method 3: Ensure MTU Consistency**
+
+Make sure both endpoints use the same MTU:
+
+```bash
+# Check device MTU
+ibstatus mlx5_0
+```
+
+---
+
+## Multi-Node Testing
+
+For multi-node tests, producer and consumer run on different machines connected via InfiniBand switch.
+
+### Prerequisites
+
+1. Both machines have Mooncake and RDMA drivers installed
+2. Both machines are in the same InfiniBand subnet
+3. Switch is properly configured
+
+### Configuration
+
+**Machine A (Producer):**
+
+```bash
+# Set RDMA host IP (InfiniBand interface IP)
+export RDMA_TEST_HOST='10.0.0.1'
+
+# Optional: Specify device
+export RDMA_DEVICE_NAME='mlx5_0'
+```
+
+**Machine B (Consumer):**
+
+```bash
+# Set RDMA host IP
+export RDMA_TEST_HOST='10.0.0.2'
+
+# Optional: Specify device
+export RDMA_DEVICE_NAME='mlx5_0'
+```
+
+### Verify Connectivity
+
+```bash
+# Ping IB interface
+ping 10.0.0.2
+
+# Test RDMA connectivity with ibping
+# On Machine B (server)
+ibping -S
+
+# On Machine A (client)
+ibping -G <Machine_B_GID>
+```
+
+---
+
+## Running Tests
+
+### Run All RDMA Tests (Single-Node, fast suite)
+
+Slow tests (large payloads, stress, concurrency integrity) are marked `@pytest.mark.slow`. Use `-m "not slow"` to skip them in quick CI or local fast iteration.
+
+```bash
+cd tests/distributed/omni_connectors
+
+# Fast suite only (excludes slow/stress tests)
+pytest test_mooncake_transfer_engine_rdma.py test_mooncake_transfer_engine_buffer.py -v -s -m "not slow"
+```
+
+### Run Including Slow Tests
+
+```bash
+# Run ALL tests including slow/stress tests
+pytest test_mooncake_transfer_engine_rdma.py test_mooncake_transfer_engine_buffer.py -v -s
+
+# Run ONLY the slow/stress tests
+pytest test_mooncake_transfer_engine_rdma.py test_mooncake_transfer_engine_buffer.py -v -s -m slow
+```
+
+### Run Buffer Management Tests
+
+```bash
+# Fast only
+pytest test_mooncake_transfer_engine_buffer.py -v -s -m "not slow"
+
+# Including allocator invariant tests (double-free, overlap, merge)
+pytest test_mooncake_transfer_engine_buffer.py -v -s
+```
+
+### Run Specific Test Classes
+
+```bash
+# Basic connector tests
+pytest test_mooncake_transfer_engine_rdma.py::TestBasicConnector -v -s
+
+# End-to-end RDMA transfer tests
+pytest test_mooncake_transfer_engine_rdma.py::TestEndToEnd -v -s
+
+# Lifecycle & resource management tests
+pytest test_mooncake_transfer_engine_rdma.py::TestLifecycle -v -s
+
+# GPU memory pool tests (requires CUDA)
+pytest test_mooncake_transfer_engine_rdma.py::TestGPUPool -v -s
+
+# Stress / correctness tests (slow)
+pytest test_mooncake_transfer_engine_rdma.py::TestStressCorrectness -v -s
+```
+
+### RDMA Environment Diagnostics
+
+For quick diagnostics (device status, Mooncake availability, env vars, etc.),
+see the [Troubleshooting section](../../../docs/design/feature/omni_connectors/mooncake_transfer_engine_connector.md#troubleshooting)
+in the connector documentation.
+
+---
+
+## Cross-Node Testing
+
+The `cross_node_mooncake_transfer_engine.py` script enables testing RDMA transfers between two separate physical machines. This script is **not** auto-discovered by `pytest` (it does not start with `test_`) — it must be run manually on each node.
+
+### Prerequisites
+
+1. Both machines have Mooncake installed
+2. Both machines are connected via InfiniBand/RoCE switch
+3. Firewall allows ZMQ ports (default: 15500, 15501)
+4. Same RDMA device name on both nodes (if multiple devices exist)
+
+### Running Cross-Node Tests
+
+**On Machine A (Producer) — start first:**
+
+```bash
+cd benchmarks/distributed/omni_connectors/
+
+# Optional: specify device if multiple exist
+export RDMA_DEVICE_NAME='mlx5_0'
+
+python cross_node_mooncake_transfer_engine.py \
+    --role producer \
+    --local-host <PRODUCER_IP> \
+    --remote-host <CONSUMER_IP> \
+    --tensor-size-mb 100 \
+    --num-transfers 3
+```
+
+**On Machine B (Consumer) — start after producer:**
+
+```bash
+cd benchmarks/distributed/omni_connectors/
+
+export RDMA_DEVICE_NAME='mlx5_0'
+
+python cross_node_mooncake_transfer_engine.py \
+    --role consumer \
+    --local-host <CONSUMER_IP> \
+    --remote-host <PRODUCER_IP> \
+    --tensor-size-mb 100 \
+    --num-transfers 3
+```
+
+### Transfer Modes
+
+| Mode | Description | Example |
+|------|-------------|---------|
+| `copy` | Normal path — tensor copied to RDMA pool (default) | `--mode copy` |
+| `zerocopy` | Zero-copy path — data created directly in RDMA pool | `--mode zerocopy` |
+| `gpu` | GPU transfer — RDMA pool on GPU, uses GPUDirect | `--mode gpu --gpu-id 0` |
+
+### Benchmark Mode
+
+Skip MD5 verification and measure pure RDMA throughput:
+
+```bash
+# Producer
+python cross_node_mooncake_transfer_engine.py \
+    --role producer \
+    --local-host <PRODUCER_IP> \
+    --remote-host <CONSUMER_IP> \
+    --tensor-size-mb 1024 \
+    --num-transfers 20 \
+    --benchmark
+
+# Consumer
+python cross_node_mooncake_transfer_engine.py \
+    --role consumer \
+    --local-host <CONSUMER_IP> \
+    --remote-host <PRODUCER_IP> \
+    --tensor-size-mb 1024 \
+    --num-transfers 20 \
+    --benchmark
+```
+
+### Cross-Node Test Options
+
+| Option | Description | Default |
+|--------|-------------|---------|
+| `--role` | `producer` or `consumer` | Required |
+| `--local-host` | Local RDMA IP address | Required |
+| `--remote-host` | Remote RDMA IP address | Required |
+| `--local-port` | Local ZMQ port for RDMA data | 15500 |
+| `--remote-port` | Remote ZMQ port for RDMA data | 15500 |
+| `--ctrl-port` | Control channel port | 15501 |
+| `--tensor-size-mb` | Tensor size in MB | 100 |
+| `--num-transfers` | Number of transfers | 3 |
+| `--mode` | `copy`, `zerocopy`, or `gpu` | `copy` |
+| `--gpu-id` | GPU ID for GPU mode | 0 |
+| `--benchmark` | Skip MD5, pure performance test | off |
+
+---
+
+## Troubleshooting
+
+### 1. "Failed to modify QP to RTR" Error
+
+**Cause**: QP handshake failed, usually due to device configuration mismatch.
+
+**Solution**:
+```bash
+# Force using the same device
+export RDMA_DEVICE_NAME='mlx5_0'
+```
+
+### 2. "Mooncake TransferEngine is not available"
+
+**Cause**: Mooncake not installed or import failed.
+
+**Solution**:
+```bash
+# Check Mooncake installation
+python -c "from mooncake.engine import TransferEngine; print('OK')"
+
+# Reinstall if needed
+pip install mooncake-transfer-engine
+# Or using uv
+uv pip install mooncake-transfer-engine
+
+```
+
+### 3. "Permission denied" accessing /dev/infiniband
+
+**Cause**: Container lacks IB device access permissions.
+
+**Solution**:
+```bash
+docker run --device=/dev/infiniband --cap-add=IPC_LOCK ...
+```
+
+### 4. Test Timeout
+
+**Cause**: RDMA connection establishment failed or network latency.
+
+**Solution**:
+```bash
+# Check network status
+ibstat
+ibstatus
+```
+
+### 5. GPU Test Failed "CUDA is not available"
+
+**Cause**: CUDA environment not configured or GPU unavailable.
+
+**Solution**:
+```bash
+# Check CUDA
+python -c "import torch; print(torch.cuda.is_available())"
+
+# Docker needs NVIDIA runtime
+docker run --gpus all ...
+```
+
+---
+
+## Environment Variables Reference
+
+| Variable | Description | Example |
+|----------|-------------|---------|
+| `RDMA_DEVICE_NAME` | Specify RDMA device name | `mlx5_0` |
+| `RDMA_TEST_HOST` | Specify test host IP | `10.0.0.1` |
+| `MC_TE_METRIC` | Enable Mooncake metrics | `1` |
+| `MC_IB_PCI_RELAXED_ORDERING` | Enable PCIe relaxed ordering | `1` |
+
+---
+
+## Test Files Overview
+
+| File | Description | Auto-discovered by pytest |
+|------|-------------|--------------------------|
+| `test_mooncake_transfer_engine_rdma.py` | Integration tests for MooncakeTransferEngineConnector (basic, E2E, lifecycle, GPU) | Yes |
+| `test_mooncake_transfer_engine_buffer.py` | Memory pool and buffer management unit tests | Yes |
+| `cross_node_mooncake_transfer_engine.py` | Cross-node (multi-machine) testing script — run manually | No (filename does not start with `test_`) |
+
+### test_mooncake_transfer_engine_rdma.py — Test Classes
+
+| Test Class | Memory Pool | Marker | Description |
+|------------|-------------|--------|-------------|
+| `TestBasicConnector` | CPU | — | Initialization, put tensor/bytes/object, cleanup, pool exhaustion |
+| `TestEndToEnd` | CPU | — | E2E RDMA transfer: tensor, bytes, object, zero-copy, large payload (100MB), mixed types, concurrency |
+| `TestLifecycle` | CPU | — | Close, context manager, double-close safety |
+| `TestGPUPool` | GPU | — | GPU pool init, put CPU/GPU tensor, GPU E2E transfer |
+| `TestStressCorrectness` | CPU | `slow` | Concurrent put+get with MD5 integrity, bidirectional concurrency, edge cases (1-element tensor, empty bytes), 500MB payload, rapid alloc/free cycles |
+
+### test_mooncake_transfer_engine_buffer.py — Test Classes
+
+| Test Class | Marker | Description |
+|------------|--------|-------------|
+| `TestBufferAllocator` | — | Basic alloc/free, alignment, exhaustion/recovery, thread safety |
+| `TestAllocatorInvariants` | `slow` | Double-free safety, overlap corruption detection, adjacent-block merging, fragmentation/defrag |
+| `TestManagedBuffer` | — | Tensor views, context manager |

benchmarks/distributed/omni_connectors/cross_node_mooncake_transfer_engine.py

@@ -0,0 +1,644 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Cross-Node RDMA Test Script (Automated Version)
+
+This script enables testing RDMA transfers between two separate machines.
+Supports three transfer modes:
+  - copy:      Normal path - tensor copied to RDMA pool (default)
+  - zerocopy:  Zero-copy path - data created directly in RDMA pool
+  - gpu:       GPU transfer - RDMA pool on GPU, uses GPUDirect
+
+Usage:
+    # On Machine A (Producer) - start first:
+    python cross_node_mooncake_transfer_engine.py --role producer --local-host hostname_A --remote-host hostname_B
+
+    # On Machine B (Consumer) - start after producer:
+    python cross_node_mooncake_transfer_engine.py --role consumer --local-host hostname_B --remote-host hostname_A
+
+    # Zero-copy mode:
+    python cross_node_mooncake_transfer_engine.py --role producer ... --mode zerocopy
+
+    # GPU mode (requires GPUDirect RDMA support):
+    python cross_node_mooncake_transfer_engine.py --role producer ... --mode gpu --gpu-id 0
+
+    # Benchmark mode (skip random data generation and MD5 verification,
+    # measures pure RDMA throughput):
+    python cross_node_mooncake_transfer_engine.py --role producer ... --benchmark
+
+Environment Variables:
+    RDMA_DEVICE_NAME:              Specify RDMA device (e.g., mlx5_0)
+    MC_IB_PCI_RELAXED_ORDERING:    Set to 1 to enable PCIe relaxed ordering
+                                   for higher RDMA throughput
+"""
+
+import argparse
+import hashlib
+import os
+import sys
+import time
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Any
+
+import msgspec
+import torch
+import zmq
+
+# Add parent path for imports
+sys.path.insert(
+    0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))))
+)
+
+from vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector import (
+    ManagedBuffer,
+    MooncakeTransferEngineConnector,
+    TransferEngine,
+)
+
+
+def compute_md5(tensor: torch.Tensor) -> str:
+    """Compute MD5 checksum of a tensor."""
+    if tensor.is_cuda:
+        tensor = tensor.cpu()
+    data = tensor.contiguous().view(torch.uint8).numpy().tobytes()
+    return hashlib.md5(data).hexdigest()
+
+
+# Control channel message types
+class CtrlMsg(msgspec.Struct):
+    """Control channel message."""
+
+    msg_type: str  # "READY", "TRANSFER", "ACK", "DONE", "ERROR"
+    request_id: str = ""
+    md5: str = ""
+    data_size: int = 0
+    error: str = ""
+
+
+@dataclass
+class TransferConfig:
+    """Configuration for cross-node transfer test."""
+
+    local_host: str
+    remote_host: str
+    local_port: int
+    remote_port: int
+    ctrl_port: int
+    num_transfers: int
+    tensor_size_mb: int
+    mode: str  # "copy", "zerocopy", "gpu"
+    gpu_id: int = 0
+    benchmark: bool = False  # Skip MD5 verification for pure performance test
+
+
+@dataclass
+class TransferStats:
+    """Statistics for transfer operations."""
+
+    success_count: int = 0
+    fail_count: int = 0
+    total_bytes: int = 0
+    elapsed_time: float = 0.0
+
+    @property
+    def throughput_mbps(self) -> float:
+        if self.elapsed_time > 0:
+            return (self.total_bytes / (1024 * 1024)) / self.elapsed_time
+        return 0.0
+
+    def print_summary(self, role: str):
+        print(f"\n{'=' * 60}")
+        print(f" {role.upper()} SUMMARY")
+        print(f"  Successful: {self.success_count}/{self.success_count + self.fail_count}")
+        print(f"  Failed:     {self.fail_count}/{self.success_count + self.fail_count}")
+        print(f"  Total:      {self.total_bytes / (1024 * 1024):.2f} MB")
+        print(f"  Time:       {self.elapsed_time:.2f} s")
+        print(f"  Throughput: {self.throughput_mbps:.2f} MB/s")
+        print(f"{'=' * 60}")
+
+
+class CrossNodeTester(ABC):
+    """Abstract base class for cross-node RDMA testing."""
+
+    def __init__(self, config: TransferConfig):
+        self.config = config
+        self.connector: MooncakeTransferEngineConnector | None = None
+        self.zmq_ctx: zmq.Context | None = None
+        self.ctrl_socket: zmq.Socket | None = None
+        self.stats = TransferStats()
+
+    def get_connector_config(self) -> dict:
+        """Get connector configuration based on mode."""
+        pool_size = int(self.config.tensor_size_mb * 1.5) * 1024 * 1024
+        pool_size = max(pool_size, 128 * 1024 * 1024)
+
+        conn_config = {
+            "host": self.config.local_host,
+            "zmq_port": self.config.local_port,
+            "protocol": "rdma",
+            "memory_pool_size": pool_size,
+        }
+
+        # Set device based on mode
+        if self.config.mode == "gpu":
+            conn_config["memory_pool_device"] = f"cuda:{self.config.gpu_id}"
+        else:
+            conn_config["memory_pool_device"] = "cpu"
+
+        # RDMA device name from environment
+        device_name = os.environ.get("RDMA_DEVICE_NAME")
+        if device_name:
+            conn_config["device_name"] = device_name
+            print(f"[CONFIG] Using RDMA device: {device_name}")
+
+        return conn_config
+
+    def initialize(self):
+        """Initialize connector and ZMQ context."""
+        print(f"[{self.role}] Initializing connector...")
+        conn_config = self.get_connector_config()
+        self.connector = MooncakeTransferEngineConnector(conn_config)
+        self.zmq_ctx = zmq.Context()
+        print(f"[{self.role}] Ready at {self.config.local_host}:{self.config.local_port}")
+
+    def cleanup(self):
+        """Cleanup resources."""
+        if self.ctrl_socket:
+            self.ctrl_socket.close()
+        if self.zmq_ctx:
+            self.zmq_ctx.term()
+        if self.connector:
+            self.connector.close()
+        print(f"[{self.role}] Closed.")
+
+    @property
+    @abstractmethod
+    def role(self) -> str:
+        pass
+
+    @abstractmethod
+    def run(self):
+        pass
+
+
+class Producer(CrossNodeTester):
+    """Producer node - sends data to consumer."""
+
+    @property
+    def role(self) -> str:
+        return "PRODUCER"
+
+    def print_header(self):
+        print(f"\n{'=' * 60}")
+        print(f" PRODUCER MODE ({self.config.mode.upper()})")
+        print(f" Local:  {self.config.local_host}:{self.config.local_port}")
+        print(f" Remote: {self.config.remote_host}:{self.config.remote_port}")
+        print(f" Control Port: {self.config.ctrl_port}")
+        print(f" Transfer Mode: {self.config.mode}")
+        if self.config.mode == "gpu":
+            print(f" GPU ID: {self.config.gpu_id}")
+        print(f"{'=' * 60}\n")
+
+    def setup_control_channel(self):
+        """Setup ZMQ control channel as server (REP socket)."""
+        self.ctrl_socket = self.zmq_ctx.socket(zmq.REP)
+        self.ctrl_socket.bind(f"tcp://*:{self.config.ctrl_port}")
+        print(f"[PRODUCER] Control channel listening on port {self.config.ctrl_port}")
+        print("[PRODUCER] Waiting for consumer to connect...")
+
+    def wait_for_consumer(self) -> bool:
+        """Wait for consumer READY signal."""
+        msg_data = self.ctrl_socket.recv()
+        msg = msgspec.msgpack.decode(msg_data, type=CtrlMsg)
+        if msg.msg_type != "READY":
+            print(f"[PRODUCER] Unexpected message: {msg.msg_type}")
+            return False
+        print("[PRODUCER] Consumer connected!")
+        self.ctrl_socket.send(msgspec.msgpack.encode(CtrlMsg(msg_type="ACK")))
+        return True
+
+    def create_test_data(self, transfer_idx: int) -> tuple[Any, str, int]:
+        """
+        Create test data based on transfer mode.
+
+        Returns:
+            (data, md5, size) tuple
+        """
+        num_elements = (self.config.tensor_size_mb * 1024 * 1024) // 4
+        data_size = num_elements * 4
+
+        if self.config.mode == "zerocopy":
+            # Zero-Copy Path: Allocate directly from connector's pool
+            offset = self.connector.allocator.alloc(data_size)
+            managed_buf = ManagedBuffer(self.connector.allocator, offset, data_size, self.connector.pool)
+
+            if self.config.benchmark:
+                # In benchmark mode, skip random data generation (use uninitialized memory)
+                return managed_buf, "", data_size
+            else:
+                # Fill buffer with random data using tensor view
+                tensor_view = managed_buf.as_tensor(dtype=torch.float32, shape=(num_elements,))
+                random_data = torch.randn(num_elements, dtype=torch.float32)
+                if tensor_view.is_cuda:
+                    tensor_view.copy_(random_data.to(tensor_view.device))
+                else:
+                    tensor_view.copy_(random_data)
+                md5 = compute_md5(tensor_view)
+                return managed_buf, md5, data_size
+
+        elif self.config.mode == "gpu":
+            # GPU Path: Create tensor on GPU
+            device = f"cuda:{self.config.gpu_id}"
+            if self.config.benchmark:
+                # In benchmark mode, use empty tensor (no random generation)
+                gpu_tensor = torch.empty(num_elements, dtype=torch.float32, device=device)
+                return gpu_tensor, "", data_size
+            else:
+                cpu_tensor = torch.randn(num_elements, dtype=torch.float32)
+                md5 = compute_md5(cpu_tensor)
+                gpu_tensor = cpu_tensor.to(device)
+                return gpu_tensor, md5, data_size
+
+        else:
+            # Copy Path (default): Create regular CPU tensor
+            if self.config.benchmark:
+                # In benchmark mode, use empty tensor (no random generation)
+                tensor = torch.empty(num_elements, dtype=torch.float32)
+                return tensor, "", data_size
+            else:
+                tensor = torch.randn(num_elements, dtype=torch.float32)
+                md5 = compute_md5(tensor)
+                return tensor, md5, data_size
+
+    def do_transfer(self, transfer_idx: int) -> bool:
+        """Perform a single transfer."""
+        req_id = f"cross_node_transfer_{transfer_idx}"
+
+        if not self.config.benchmark:
+            print(f"\n[PRODUCER] Transfer {transfer_idx + 1}/{self.config.num_transfers}")
+
+        # Create test data
+        t0 = time.time()
+        data, md5, data_size = self.create_test_data(transfer_idx)
+        t_create = time.time() - t0
+
+        if not self.config.benchmark:
+            print(f"  Mode: {self.config.mode}")
+            print(f"  Size: {self.config.tensor_size_mb} MB")
+            if md5:
+                print(f"  MD5:  {md5[:16]}...")
+            print(f"  Create time: {t_create * 1000:.1f} ms")
+
+        # Put data
+        t1 = time.time()
+        success, size, metadata = self.connector.put("producer", "consumer", req_id, data)
+        t_put = time.time() - t1
+
+        if not success:
+            print("  [FAIL] Put failed")
+            return False
+
+        if not self.config.benchmark:
+            print(f"  [OK] Put successful, {size} bytes ({t_put * 1000:.1f} ms)")
+
+        # Wait for consumer to request transfer info
+        msg_data = self.ctrl_socket.recv()
+        msg = msgspec.msgpack.decode(msg_data, type=CtrlMsg)
+
+        if msg.msg_type != "READY":
+            print(f"  [ERROR] Unexpected message: {msg.msg_type}")
+            return False
+
+        # Send transfer metadata to consumer
+        transfer_msg = CtrlMsg(
+            msg_type="TRANSFER",
+            request_id=req_id,
+            md5=md5,
+            data_size=data_size,
+        )
+        self.ctrl_socket.send(msgspec.msgpack.encode(transfer_msg))
+
+        # Wait for consumer ACK (this includes RDMA transfer time)
+        t2 = time.time()
+        msg_data = self.ctrl_socket.recv()
+        t_rdma = time.time() - t2
+        msg = msgspec.msgpack.decode(msg_data, type=CtrlMsg)
+
+        success = msg.msg_type == "ACK"
+        if success:
+            if not self.config.benchmark:
+                print(f"  [OK] RDMA transfer complete ({t_rdma * 1000:.1f} ms)")
+            self.stats.success_count += 1
+            self.stats.total_bytes += size
+        else:
+            print(f"  [WARN] Consumer reported error: {msg.error}")
+            self.stats.fail_count += 1
+
+        # Send ACK to allow consumer to continue
+        self.ctrl_socket.send(msgspec.msgpack.encode(CtrlMsg(msg_type="ACK")))
+
+        # Cleanup buffer
+        self.connector.cleanup(req_id)
+
+        return success
+
+    def run(self):
+        """Run the producer."""
+        self.print_header()
+        self.initialize()
+        self.setup_control_channel()
+
+        try:
+            if not self.wait_for_consumer():
+                return
+
+            if self.config.benchmark:
+                print(
+                    f"[BENCHMARK] Running {self.config.num_transfers} "
+                    f"transfers of {self.config.tensor_size_mb} MB each..."
+                )
+
+            start_time = time.time()
+
+            for i in range(self.config.num_transfers):
+                self.do_transfer(i)
+                if self.config.benchmark and (i + 1) % 10 == 0:
+                    elapsed = time.time() - start_time
+                    current_throughput = (self.stats.total_bytes / (1024 * 1024)) / elapsed
+                    print(f"  Progress: {i + 1}/{self.config.num_transfers}, Throughput: {current_throughput:.2f} MB/s")
+
+            self.stats.elapsed_time = time.time() - start_time
+            self.stats.print_summary("PRODUCER")
+
+            # Wait for final consumer message and send DONE
+            self.ctrl_socket.recv()
+            self.ctrl_socket.send(msgspec.msgpack.encode(CtrlMsg(msg_type="DONE")))
+
+        finally:
+            self.cleanup()
+
+
+class Consumer(CrossNodeTester):
+    """Consumer node - receives data from producer."""
+
+    @property
+    def role(self) -> str:
+        return "CONSUMER"
+
+    def print_header(self):
+        print(f"\n{'=' * 60}")
+        print(f" CONSUMER MODE ({self.config.mode.upper()})")
+        print(f" Local:  {self.config.local_host}:{self.config.local_port}")
+        print(f" Remote: {self.config.remote_host}:{self.config.remote_port}")
+        print(f" Control Port: {self.config.ctrl_port}")
+        print(f" Transfer Mode: {self.config.mode}")
+        if self.config.mode == "gpu":
+            print(f" GPU ID: {self.config.gpu_id}")
+        print(f"{'=' * 60}\n")
+
+    def setup_control_channel(self):
+        """Setup ZMQ control channel as client (REQ socket)."""
+        self.ctrl_socket = self.zmq_ctx.socket(zmq.REQ)
+        ctrl_addr = f"tcp://{self.config.remote_host}:{self.config.ctrl_port}"
+        print(f"[CONSUMER] Connecting to producer control channel at {ctrl_addr}...")
+        self.ctrl_socket.connect(ctrl_addr)
+
+    def connect_to_producer(self) -> bool:
+        """Connect to producer and send READY signal."""
+        self.ctrl_socket.send(msgspec.msgpack.encode(CtrlMsg(msg_type="READY")))
+        msg_data = self.ctrl_socket.recv()
+        msg = msgspec.msgpack.decode(msg_data, type=CtrlMsg)
+        if msg.msg_type != "ACK":
+            print(f"[CONSUMER] Unexpected response: {msg.msg_type}")
+            return False
+        print("[CONSUMER] Connected to producer! Starting transfers...")
+        return True
+
+    def do_transfer(self, transfer_idx: int) -> bool:
+        """Perform a single transfer."""
+        if not self.config.benchmark:
+            print(f"\n[CONSUMER] Transfer {transfer_idx + 1}/{self.config.num_transfers}")
+
+        # Request next transfer info
+        self.ctrl_socket.send(msgspec.msgpack.encode(CtrlMsg(msg_type="READY")))
+        msg_data = self.ctrl_socket.recv()
+        msg = msgspec.msgpack.decode(msg_data, type=CtrlMsg)
+
+        if msg.msg_type == "DONE":
+            print("[CONSUMER] Producer signaled completion")
+            return False
+
+        if msg.msg_type != "TRANSFER":
+            print(f"[CONSUMER] Unexpected message: {msg.msg_type}")
+            return False
+
+        req_id = msg.request_id
+        expected_md5 = msg.md5
+        data_size = msg.data_size
+        num_elements = data_size // 4
+
+        if not self.config.benchmark:
+            print(f"  Request ID: {req_id}")
+            if expected_md5:
+                print(f"  Expected MD5: {expected_md5[:16]}...")
+            print(f"  Data Size: {data_size / (1024 * 1024):.2f} MB")
+
+        # Build metadata for get
+        metadata = {
+            "request_id": req_id,
+            "source_host": self.config.remote_host,
+            "source_port": self.config.remote_port,
+            "data_size": data_size,
+            "dtype": "float32",
+            "shape": [num_elements],
+            "is_fast_path": True,
+        }
+
+        if not self.config.benchmark:
+            print(f"  [INFO] Requesting from {self.config.remote_host}:{self.config.remote_port}")
+
+        # Get data with timing
+        t0 = time.time()
+        result = self.connector.get("producer", "consumer", req_id, metadata)
+        t_get = time.time() - t0
+
+        response_msg = CtrlMsg(msg_type="ERROR", error="Get failed")
+
+        if result is not None:
+            recv_buffer, recv_size = result
+            if not self.config.benchmark:
+                print(f"  [OK] Get successful, {recv_size} bytes ({t_get * 1000:.1f} ms)")
+
+            if isinstance(recv_buffer, ManagedBuffer):
+                # In benchmark mode, skip MD5 verification
+                if self.config.benchmark or not expected_md5:
+                    response_msg = CtrlMsg(msg_type="ACK")
+                    self.stats.success_count += 1
+                    self.stats.total_bytes += recv_size
+                else:
+                    # Verify data
+                    t1 = time.time()
+                    reconstructed = recv_buffer.as_tensor(dtype=torch.float32, shape=(num_elements,))
+                    recv_md5 = compute_md5(reconstructed)
+                    t_md5 = time.time() - t1
+                    print(f"  MD5: {recv_md5[:16]}... ({t_md5 * 1000:.1f} ms)")
+
+                    if recv_md5 == expected_md5:
+                        print("  [PASS] MD5 checksum verified!")
+                        response_msg = CtrlMsg(msg_type="ACK")
+                        self.stats.success_count += 1
+                        self.stats.total_bytes += recv_size
+                    else:
+                        print("  [FAIL] MD5 mismatch!")
+                        response_msg = CtrlMsg(msg_type="ERROR", error="MD5 mismatch")
+                        self.stats.fail_count += 1
+
+                recv_buffer.release()
+            else:
+                response_msg = CtrlMsg(msg_type="ACK")
+                self.stats.success_count += 1
+                self.stats.total_bytes += recv_size
+        else:
+            print("  [FAIL] Get failed")
+            self.stats.fail_count += 1
+
+        # Send response to producer
+        self.ctrl_socket.send(msgspec.msgpack.encode(response_msg))
+        # Wait for ACK
+        self.ctrl_socket.recv()
+
+        return response_msg.msg_type == "ACK"
+
+    def run(self):
+        """Run the consumer."""
+        self.print_header()
+        self.initialize()
+        self.setup_control_channel()
+
+        try:
+            if not self.connect_to_producer():
+                return
+
+            if self.config.benchmark:
+                print(
+                    f"[BENCHMARK] Running {self.config.num_transfers} "
+                    f"transfers of {self.config.tensor_size_mb} MB each..."
+                )
+
+            start_time = time.time()
+
+            for i in range(self.config.num_transfers):
+                if not self.do_transfer(i):
+                    break
+                if self.config.benchmark and (i + 1) % 10 == 0:
+                    elapsed = time.time() - start_time
+                    current_throughput = (self.stats.total_bytes / (1024 * 1024)) / elapsed
+                    print(f"  Progress: {i + 1}/{self.config.num_transfers}, Throughput: {current_throughput:.2f} MB/s")
+
+            self.stats.elapsed_time = time.time() - start_time
+            self.stats.print_summary("CONSUMER")
+
+            # Send final READY and wait for DONE
+            self.ctrl_socket.send(msgspec.msgpack.encode(CtrlMsg(msg_type="READY")))
+            self.ctrl_socket.recv()
+
+        finally:
+            self.cleanup()
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Cross-Node RDMA Test (Automated)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Transfer Modes:
+  copy      - Normal path: tensor copied to RDMA pool (default)
+  zerocopy  - Zero-copy path: data created directly in RDMA pool
+  gpu       - GPU transfer: RDMA pool on GPU, uses GPUDirect
+
+Examples:
+  # Copy mode (default):
+  python cross_node_mooncake_transfer_engine.py --role producer \
+  --local-host hostA --remote-host hostB
+
+  # Zero-copy mode:
+  python cross_node_mooncake_transfer_engine.py --role producer \
+  --local-host hostA --remote-host hostB --mode zerocopy
+
+  # GPU mode:
+  python cross_node_mooncake_transfer_engine.py --role producer \
+  --local-host hostA --remote-host hostB --mode gpu --gpu-id 0
+
+  # Benchmark mode (skip MD5, measure pure RDMA performance):
+  python cross_node_mooncake_transfer_engine.py --role producer \
+  --local-host hostA --remote-host hostB --benchmark
+
+  # With specific RDMA device:
+  RDMA_DEVICE_NAME=mlx5_0 python cross_node_mooncake_transfer_engine.py --role producer ...
+        """,
+    )
+
+    parser.add_argument(
+        "--role", required=True, choices=["producer", "consumer"], help="Role: producer (sends) or consumer (receives)"
+    )
+    parser.add_argument("--local-host", required=True, help="Local hostname or IP address")
+    parser.add_argument("--remote-host", required=True, help="Remote hostname or IP address")
+    parser.add_argument("--local-port", type=int, default=15500, help="Local ZMQ port for RDMA data (default: 15500)")
+    parser.add_argument("--remote-port", type=int, default=15500, help="Remote ZMQ port for RDMA data (default: 15500)")
+    parser.add_argument("--ctrl-port", type=int, default=15501, help="Control channel port (default: 15501)")
+    parser.add_argument("--num-transfers", type=int, default=20, help="Number of transfers to perform (default: 3)")
+    parser.add_argument("--tensor-size-mb", type=int, default=100, help="Tensor size in MB (default: 100)")
+    parser.add_argument(
+        "--mode",
+        choices=["copy", "zerocopy", "gpu"],
+        default="copy",
+        help="Transfer mode: copy, zerocopy, or gpu (default: copy)",
+    )
+    parser.add_argument("--gpu-id", type=int, default=0, help="GPU ID for GPU mode (default: 0)")
+    parser.add_argument(
+        "--benchmark", action="store_true", help="Benchmark mode: skip MD5 verification for pure performance test"
+    )
+
+    args = parser.parse_args()
+
+    # Check Mooncake
+    if TransferEngine is None:
+        print("[ERROR] Mooncake TransferEngine is not available.")
+        print("Install with: pip install mooncake")
+        sys.exit(1)
+
+    # Check CUDA for GPU mode
+    if args.mode == "gpu":
+        if not torch.cuda.is_available():
+            print("[ERROR] CUDA is not available but GPU mode was requested.")
+            sys.exit(1)
+        if args.gpu_id >= torch.cuda.device_count():
+            print(f"[ERROR] GPU {args.gpu_id} not available. Found {torch.cuda.device_count()} GPUs.")
+            sys.exit(1)
+        print(f"[INFO] Using GPU {args.gpu_id}: {torch.cuda.get_device_name(args.gpu_id)}")
+
+    config = TransferConfig(
+        local_host=args.local_host,
+        remote_host=args.remote_host,
+        local_port=args.local_port,
+        remote_port=args.remote_port,
+        ctrl_port=args.ctrl_port,
+        num_transfers=args.num_transfers,
+        tensor_size_mb=args.tensor_size_mb,
+        mode=args.mode,
+        gpu_id=args.gpu_id,
+        benchmark=args.benchmark,
+    )
+
+    if args.role == "producer":
+        tester = Producer(config)
+    else:
+        tester = Consumer(config)
+
+    tester.run()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/qwen3-omni/README.md

@@ -36,7 +36,7 @@ bash benchmarks/qwen3-omni/transformers/eval_qwen3_moe_omni_transformers.sh
 ```
 
 What it does:
-- Runs `qwen3_omni_moe_transformers.py` over `top100.txt` with `--num_prompts 100`.
+- Runs `qwen3_omni_moe_transformers.py` over `top100.txt` with `--num-prompts 100`.
 - Outputs to `benchmarks/qwen3-omni/transformers/benchmark_results/`:
   - `perf_stats.json` — aggregated & per-prompt TPS/latency (thinker/talker/code2wav/overall).
   - `results.json` — per-prompt outputs and audio paths.
@@ -53,7 +53,7 @@ bash benchmarks/qwen3-omni/vllm_omni/eval_qwen3_moe_omni.sh
 ```
 
 What it does:
-- Runs `examples/offline_inference/qwen3_omni/end2end.py` with `--enable-stats`.
+- Runs `examples/offline_inference/qwen3_omni/end2end.py` with `--log-stats`.
 - Uses `benchmarks/build_dataset/top100.txt` and writes to:
   - Logs: `benchmarks/qwen3-omni/vllm_omni/logs/`
     - `omni_llm_pipeline_text.orchestrator.stats.jsonl` — per-stage latency stats.

benchmarks/qwen3-omni/transformers/eval_qwen3_moe_omni_transformers.sh

@@ -16,7 +16,7 @@ if [[ ! -f "benchmarks/qwen3-omni/transformers/qwen3_omni_moe_transformers.py" ]
 else
     cd benchmarks/qwen3-omni/transformers
 
-    python qwen3_omni_moe_transformers.py --prompts_file ../../build_dataset/top100.txt --num_prompts 100
+    python qwen3_omni_moe_transformers.py --prompts-file ../../build_dataset/top100.txt --num-prompts 100
 
     echo "Logs and outputs are saved to $(pwd)/benchmark_results:"
     echo "  - perf_stats.json    Aggregated/per-prompt TPS and latency (thinker/talker/code2wav/overall)"

benchmarks/qwen3-omni/transformers/qwen3_omni_moe_transformers.py

@@ -227,17 +227,17 @@ def print_stats(stats: dict):
 def main():
     parser = argparse.ArgumentParser(description="Qwen3-Omni Benchmark Script")
     parser.add_argument(
-        "--prompts_file",
+        "--prompts-file",
         type=str,
         default="benchmark/build_dataset/top100.txt",
         help="Path to the prompts file (one prompt per line)",
     )
     parser.add_argument(
-        "--output_dir", type=str, default="benchmark_results", help="Directory to save benchmark results"
+        "--output-dir", type=str, default="benchmark_results", help="Directory to save benchmark results"
     )
-    parser.add_argument("--model_path", type=str, default=MODEL_PATH, help="Path to the model")
+    parser.add_argument("--model-path", type=str, default=MODEL_PATH, help="Path to the model")
     parser.add_argument("--speaker", type=str, default="Ethan", help="Speaker voice for audio output")
-    parser.add_argument("--num_prompts", type=int, default=None, help="Number of prompts to process (default: all)")
+    parser.add_argument("--num-prompts", type=int, default=None, help="Number of prompts to process (default: all)")
     args = parser.parse_args()
 
     # Load model and processor

benchmarks/qwen3-omni/vllm_omni/eval_qwen3_moe_omni.sh

@@ -23,7 +23,7 @@ else
     python $end2end_script_path --output-wav $outputs_dir \
                       --query-type text \
                       --txt-prompts $build_dataset_path \
-                      --enable-stats \
+                      --log-stats \
                       --log-dir $log_dir
     echo "Logs and outputs are saved in ${log_dir} and ${outputs_dir} respectively:"
     echo "  - omni_llm_pipeline_text                       run dir/base name"

docker/Dockerfile.ci

@@ -1,5 +1,5 @@
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.15.0
+ARG VLLM_BASE_TAG=v0.16.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}
@@ -8,12 +8,13 @@ COPY . .
 
 # Install system dependencies
 RUN apt-get update && \
-    apt-get install -y ffmpeg && \
+    apt-get install -y ffmpeg git sox libsox-fmt-all jq && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
 # Install vllm-omni into the same uv-managed Python environment used by the base image.
-RUN uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" --no-cache-dir ".[dev]"
+# Use bash -c so that $(python3 -c ...) is expanded inside the container.
+RUN uv pip install --system --no-cache-dir ".[dev]"
 
 RUN ln -sf /usr/bin/python3 /usr/bin/python
 

docker/Dockerfile.npu

@@ -2,13 +2,16 @@ ARG VLLM_ASCEND_IMAGE=quay.io/ascend/vllm-ascend
 ARG VLLM_ASCEND_TAG=v0.14.0rc1
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
+WORKDIR /vllm-workspace/vllm-ascend
+RUN git checkout e2175d9c7e62b437391dfee996b1375674ba7c18
+RUN pip install -v -e .
+
 ARG APP_DIR=/vllm-workspace/vllm-omni
 WORKDIR ${APP_DIR}
 
 COPY . .
 
-# Install vllm-omni with dev dependencies
-RUN pip install --no-cache-dir -e . --no-build-isolation
+RUN pip install -v -e .
 
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 

docker/Dockerfile.npu.a3

@@ -2,13 +2,16 @@ ARG VLLM_ASCEND_IMAGE=quay.io/ascend/vllm-ascend
 ARG VLLM_ASCEND_TAG=v0.14.0rc1-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
+WORKDIR /vllm-workspace/vllm-ascend
+RUN git checkout e2175d9c7e62b437391dfee996b1375674ba7c18
+RUN pip install -v -e .
+
 ARG APP_DIR=/vllm-workspace/vllm-omni
 WORKDIR ${APP_DIR}
 
 COPY . .
 
-# Install vllm-omni with dev dependencies
-RUN pip install --no-cache-dir -e . --no-build-isolation
+RUN pip install -v -e .
 
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 

docker/Dockerfile.rocm

@@ -1,28 +1,53 @@
-ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.15.0
-FROM ${BASE_IMAGE} AS final
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.16.0
+FROM ${BASE_IMAGE} AS base
+
+# Declare a variable to know if we want to use the nightly build or the stable build.
+# NOTE: REMINDER to vLLM-Omni rebase maintainer
+# Remember to set `USE_NIGHTLY_BUILD` to 0, when switching back to stable vLLM docker image
+# 1. If vLLM-omni maintainer is forced to use custom commits
+#    during rebasing, they can change the two variables
+# 2. Whenever vLLM upstream has released a stable version,
+#    we should swap over to use stable release ASAP.
+#    We should avoid relying on custom commits.
+ARG USE_NIGHTLY_BUILD=0
+ARG VLLM_VERSION_OR_COMMIT_HASH=2d5be1dd5ce2e44dfea53ea03ff61143da5137eb
+ARG ARG_PYTORCH_ROCM_ARCH
+ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}}
 
 ARG COMMON_WORKDIR=/app
-
 WORKDIR ${COMMON_WORKDIR}
 
 # Step 1: Setup - Install system dependencies
 RUN apt-get update && \
-    apt-get install -y ffmpeg && \
+    apt-get install -y ffmpeg git sox libsox-fmt-all jq && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
-RUN mkdir -p ${COMMON_WORKDIR}/vllm-omni
+# Step 2: Conditionally reinstall vllm from source for nightly builds
+RUN if [ "${USE_NIGHTLY_BUILD}" = "1" ]; then \
+        python3 -m pip uninstall -y vllm && rm -rf vllm && \
+        git clone https://github.com/vllm-project/vllm.git && \
+        cd vllm && \
+        git checkout ${VLLM_VERSION_OR_COMMIT_HASH} && \
+        python3 -m pip install -r requirements/rocm.txt && \
+        python3 setup.py clean --all && \
+        PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} python3 setup.py develop && \
+        cd ../ && \
+        rm -rf vllm/.git; \
+    fi
 
-# Step 2: Copy vllm-omni code and install without uv
+# Step 3: Copy vllm-omni code and install without uv
+RUN mkdir -p ${COMMON_WORKDIR}/vllm-omni
 COPY . ${COMMON_WORKDIR}/vllm-omni
 RUN cd ${COMMON_WORKDIR}/vllm-omni && uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" --no-cache-dir ".[dev]" --no-build-isolation
 
 RUN ln -sf /usr/bin/python3 /usr/bin/python
 
-CMD ["/bin/bash"]
+FROM base AS test
 
+CMD ["/bin/bash"]
 ENTRYPOINT []
 
-#Set entrypoint for vllm-openai official images
-FROM final AS vllm-openai
+# Set entrypoint for vllm-openai official images
+FROM base AS vllm-openai
 ENTRYPOINT ["vllm", "serve", "--omni"]

docker/Dockerfile.xpu

@@ -0,0 +1,139 @@
+# Argument to configure vllm base image if pre-built
+ARG VLLM_BASE=vllm-base
+
+FROM intel/deep-learning-essentials:2025.3.2-0-devel-ubuntu24.04 AS vllm-base
+
+WORKDIR /workspace/
+
+ARG PYTHON_VERSION=3.12
+ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"
+
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
+    add-apt-repository -y ppa:kobuk-team/intel-graphics
+
+RUN apt clean && apt-get update -y && \
+    apt-get install -y --no-install-recommends --fix-missing \
+    curl \
+    ffmpeg \
+    git \
+    libsndfile1 \
+    libsm6 \
+    libxext6 \
+    libgl1 \
+    lsb-release \
+    libaio-dev \
+    numactl \
+    wget \
+    vim \
+    python3.12 \
+    python3.12-dev \
+    python3-pip
+
+RUN apt update && apt upgrade -y && \
+    apt install -y libze1 libze-dev libze-intel-gpu1 intel-opencl-icd libze-intel-gpu-raytracing intel-ocloc && \
+    apt install -y intel-oneapi-compiler-dpcpp-cpp-2025.3
+
+ENV PATH="/root/.local/bin:$PATH"
+ENV VIRTUAL_ENV="/opt/venv"
+ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+
+# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.2.
+ARG ONECCL_INSTALLER="intel-oneccl-2021.15.7.8_offline.sh"
+RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.7/${ONECCL_INSTALLER}" && \
+    bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
+    rm "${ONECCL_INSTALLER}" && \
+    echo "source /opt/intel/oneapi/setvars.sh --force" >> /root/.bashrc && \
+    echo "source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force" >> /root/.bashrc
+RUN rm -f /opt/intel/oneapi/ccl/latest && \
+    ln -s /opt/intel/oneapi/ccl/2021.15 /opt/intel/oneapi/ccl/latest
+
+SHELL ["bash", "-c"]
+CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
+
+WORKDIR /workspace/
+ENV UV_HTTP_TIMEOUT=500
+
+# Configure package index for XPU
+ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+ENV UV_LINK_MODE="copy"
+
+ARG VLLM_VERSION=v0.16.0
+RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
+WORKDIR /workspace/vllm
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --upgrade pip && \
+    uv pip install -r requirements/xpu.txt
+
+ # used for suffix method speculative decoding
+ # build deps for proto + nanobind-based extensions to set up the build environment
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install grpcio-tools protobuf nanobind
+ # arctic-inference is built from source which needs torch-xpu properly installed first
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/intel/oneapi/setvars.sh --force && \
+    source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force && \
+    export CMAKE_PREFIX_PATH="$(python -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH}" && \
+    uv pip install --no-build-isolation arctic-inference==0.1.1
+
+ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
+
+ENV VLLM_TARGET_DEVICE=xpu
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --no-build-isolation .
+
+CMD ["/bin/bash"]
+
+FROM vllm-base AS vllm-openai
+
+# install additional dependencies for openai api server
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install accelerate hf_transfer pytest pytest_asyncio lm_eval[api] modelscope
+
+# install development dependencies (for testing)
+RUN uv pip install -e tests/vllm_test_utils
+
+# install nixl from source code
+ENV NIXL_VERSION=0.7.0
+RUN python /workspace/vllm/tools/install_nixl_from_source_ubuntu.py
+
+# ensure vllm is properly installed
+RUN python -c "import vllm, inspect; print(vllm.__file__)"
+RUN uv pip show vllm
+
+CMD ["/bin/bash"]
+
+ENTRYPOINT []
+
+FROM ${VLLM_BASE} AS vllm-omni
+
+WORKDIR /workspace/vllm-omni
+COPY . .
+
+ENV VLLM_OMNI_TARGET_DEVICE=xpu
+RUN uv pip install --no-cache-dir ".[dev]" --no-build-isolation
+
+# FIX triton
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip uninstall triton triton-xpu && \
+    uv pip install triton-xpu==3.6.0
+
+# remove torch bundled oneccl to avoid conflicts
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip uninstall oneccl oneccl-devel
+
+FROM vllm-omni AS vllm-omni-openai
+
+RUN ln -sf /usr/bin/python3 /usr/bin/python
+
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+ENTRYPOINT ["vllm", "serve", "--omni"]

docs/.nav.yml

@@ -8,27 +8,32 @@ nav:
     - OpenAI-Compatible API:
       - Image Generation: serving/image_generation_api.md
       - Image Edit: serving/image_edit_api.md
+      - Text to Speech: serving/speech_api.md
   - Examples:
     - examples/README.md
     - Offline Inference:
       - BAGEL-7B-MoT: user_guide/examples/offline_inference/bagel.md
+      - GLM-Image Multistage End-to-End Inference: user_guide/examples/offline_inference/glm_image.md
       - Image-To-Image: user_guide/examples/offline_inference/image_to_image.md
       - Image-To-Video: user_guide/examples/offline_inference/image_to_video.md
-      - LoRA Inference Examples: user_guide/examples/offline_inference/lora_inference.md
+      - MiMo-Audio Offline Inference: user_guide/examples/offline_inference/mimo_audio.md
       - Qwen2.5-Omni: user_guide/examples/offline_inference/qwen2_5_omni.md
       - Qwen3-Omni: user_guide/examples/offline_inference/qwen3_omni.md
-      - Qwen3-TTS Offline Inference: user_guide/examples/offline_inference/qwen3_tts.md
+      - Qwen3-TTS: user_guide/examples/offline_inference/qwen3_tts.md
       - Text-To-Audio: user_guide/examples/offline_inference/text_to_audio.md
       - Text-To-Image: user_guide/examples/offline_inference/text_to_image.md
       - Text-To-Video: user_guide/examples/offline_inference/text_to_video.md
     - Online Serving:
       - BAGEL-7B-MoT: user_guide/examples/online_serving/bagel.md
+      - GLM-Image Online Serving: user_guide/examples/online_serving/glm_image.md
       - Image-To-Image: user_guide/examples/online_serving/image_to_image.md
-      - Online LoRA Inference (Diffusion): user_guide/examples/online_serving/lora_inference.md
+      - Image-To-Video: user_guide/examples/online_serving/image_to_video.md
+      - Online serving Example of vLLM-Omni for MiMo-Audio: user_guide/examples/online_serving/mimo_audio.md
       - Qwen2.5-Omni: user_guide/examples/online_serving/qwen2_5_omni.md
       - Qwen3-Omni: user_guide/examples/online_serving/qwen3_omni.md
-      - Qwen3-TTS Online Serving: user_guide/examples/online_serving/qwen3_tts.md
+      - Qwen3-TTS: user_guide/examples/online_serving/qwen3_tts.md
       - Text-To-Image: user_guide/examples/online_serving/text_to_image.md
+      - Text-To-Video: user_guide/examples/online_serving/text_to_video.md
   - General:
     - usage/*
   - Configuration:
@@ -39,15 +44,23 @@ nav:
   - Features:
     - Sleep Mode: features/sleep_mode.md
     - Diffusion Features:
-      - Overview: user_guide/diffusion_acceleration.md
+      - Acceleration Overview: user_guide/diffusion_acceleration.md
       - TeaCache: user_guide/diffusion/teacache.md
       - Cache-DiT: user_guide/diffusion/cache_dit_acceleration.md
+      - Quantization:
+        - Overview: user_guide/diffusion/quantization/overview.md
+        - FP8: user_guide/diffusion/quantization/fp8.md
+        - GGUF: user_guide/diffusion/quantization/gguf.md
       - Parallelism Acceleration: user_guide/diffusion/parallelism_acceleration.md
       - CPU Offloading: user_guide/diffusion/cpu_offload_diffusion.md
+      - LoRA: user_guide/diffusion/lora.md
+      - Hybrid Sharded Data Parallel: design/feature/hsdp.md
+      - Custom Pipeline: features/custom_pipeline.md
     - ComfyUI: features/comfyui.md
 - Developer Guide:
   - General:
     - contributing/README.md
+    - pr_reviewer.md
     - glob: contributing/*
       flatten_single_child_sections: true
   - Model Implementation:
@@ -62,6 +75,12 @@ nav:
       - design/feature/disaggregated_inference.md
       - design/feature/ray_based_execution.md
       - design/feature/omni_connectors/
+      - design/feature/cfg_parallel.md
+      - design/feature/sequence_parallel.md
+      - design/feature/tensor_parallel.md
+      - design/feature/cache_dit.md
+      - design/feature/teacache.md
+      - design/feature/async_chunk_design.md
     - Module Design:
       - design/module/ar_module.md
       - design/module/dit_module.md

docs/api/README.md

@@ -7,23 +7,18 @@ Main entry points for vLLM-Omni inference and serving.
 - [vllm_omni.entrypoints.async_omni.AsyncOmni][]
 - [vllm_omni.entrypoints.async_omni_diffusion.AsyncOmniDiffusion][]
 - [vllm_omni.entrypoints.async_omni_llm.AsyncOmniLLM][]
-- [vllm_omni.entrypoints.chat_utils.OmniAsyncMultiModalContentParser][]
-- [vllm_omni.entrypoints.chat_utils.OmniAsyncMultiModalItemTracker][]
-- [vllm_omni.entrypoints.chat_utils.parse_chat_messages_futures][]
 - [vllm_omni.entrypoints.cli.benchmark.base.OmniBenchmarkSubcommandBase][]
 - [vllm_omni.entrypoints.cli.benchmark.main.OmniBenchmarkSubcommand][]
 - [vllm_omni.entrypoints.cli.benchmark.serve.OmniBenchmarkServingSubcommand][]
 - [vllm_omni.entrypoints.cli.serve.OmniServeCommand][]
 - [vllm_omni.entrypoints.client_request_state.ClientRequestState][]
-- [vllm_omni.entrypoints.log_utils.OrchestratorMetrics][]
-- [vllm_omni.entrypoints.log_utils.StageRequestMetrics][]
-- [vllm_omni.entrypoints.log_utils.StageStats][]
 - [vllm_omni.entrypoints.omni.Omni][]
 - [vllm_omni.entrypoints.omni.OmniBase][]
 - [vllm_omni.entrypoints.omni_diffusion.OmniDiffusion][]
 - [vllm_omni.entrypoints.omni_llm.OmniLLM][]
 - [vllm_omni.entrypoints.omni_stage.OmniStage][]
 - [vllm_omni.entrypoints.stage_utils.OmniStageTaskType][]
+- [vllm_omni.entrypoints.zmq_utils.ZmqQueue][]
 
 ## Inputs
 
@@ -34,7 +29,6 @@ Input data structures for multi-modal inputs.
 - [vllm_omni.inputs.data.OmniTextPrompt][]
 - [vllm_omni.inputs.data.OmniTokenInputs][]
 - [vllm_omni.inputs.data.OmniTokensPrompt][]
-- [vllm_omni.inputs.parse.parse_singleton_prompt_omni][]
 - [vllm_omni.inputs.preprocess.OmniInputPreprocessor][]
 
 ## Engine
@@ -42,6 +36,12 @@ Input data structures for multi-modal inputs.
 Engine classes for offline and online inference.
 
 - [vllm_omni.diffusion.diffusion_engine.DiffusionEngine][]
+- [vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector.BufferAllocator][]
+- [vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector.ManagedBuffer][]
+- [vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector.MooncakeAgentMetadata][]
+- [vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector.MooncakeTransferEngineConnector][]
+- [vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector.QueryRequest][]
+- [vllm_omni.distributed.omni_connectors.connectors.mooncake_transfer_engine_connector.QueryResponse][]
 - [vllm_omni.engine.AdditionalInformationEntry][]
 - [vllm_omni.engine.AdditionalInformationPayload][]
 - [vllm_omni.engine.OmniEngineCoreOutput][]
@@ -78,6 +78,8 @@ 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.mimo_audio.config_mimo_audio.MiMoAudioConfig][]
+- [vllm_omni.model_executor.models.mimo_audio.config_mimo_audio.MiMoAudioTokenizerConfig][]
 - [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSConfig][]
 - [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSSpeakerEncoderConfig][]
 - [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSTalkerCodePredictorConfig][]
@@ -95,8 +97,10 @@ Configuration classes.
 Worker classes and model runners for distributed inference.
 
 - [vllm_omni.diffusion.worker.diffusion_model_runner.DiffusionModelRunner][]
+- [vllm_omni.diffusion.worker.diffusion_worker.CustomPipelineWorkerExtension][]
 - [vllm_omni.diffusion.worker.diffusion_worker.DiffusionWorker][]
 - [vllm_omni.diffusion.worker.diffusion_worker.WorkerProc][]
+- [vllm_omni.diffusion.worker.diffusion_worker.WorkerWrapperBase][]
 - [vllm_omni.platforms.npu.worker.npu_ar_model_runner.ExecuteModelState][]
 - [vllm_omni.platforms.npu.worker.npu_ar_model_runner.NPUARModelRunner][]
 - [vllm_omni.platforms.npu.worker.npu_ar_worker.NPUARWorker][]
@@ -107,10 +111,12 @@ Worker classes and model runners for distributed inference.
 - [vllm_omni.platforms.xpu.worker.xpu_ar_worker.XPUARWorker][]
 - [vllm_omni.platforms.xpu.worker.xpu_generation_model_runner.XPUGenerationModelRunner][]
 - [vllm_omni.platforms.xpu.worker.xpu_generation_worker.XPUGenerationWorker][]
+- [vllm_omni.worker.base.OmniGPUWorkerBase][]
 - [vllm_omni.worker.gpu_ar_model_runner.ExecuteModelState][]
 - [vllm_omni.worker.gpu_ar_model_runner.GPUARModelRunner][]
 - [vllm_omni.worker.gpu_ar_worker.GPUARWorker][]
 - [vllm_omni.worker.gpu_generation_model_runner.GPUGenerationModelRunner][]
 - [vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker][]
+- [vllm_omni.worker.gpu_memory_utils.parse_cuda_visible_devices][]
 - [vllm_omni.worker.gpu_model_runner.OmniGPUModelRunner][]
 - [vllm_omni.worker.mixins.OmniWorkerMixin][]

docs/cli/serve.md

@@ -0,0 +1,9 @@
+# vllm-omni serve
+
+## JSON CLI Arguments
+
+--8<-- "docs/cli/json_tip.inc.md"
+
+## Arguments
+
+--8<-- "docs/generated/argparse_omni/omni_serve.inc.md"

docs/configuration/stage_configs.md

@@ -273,3 +273,15 @@ Default: `True`
 Penalty applied to tokens that have already appeared in the generated sequence. Values greater than 1.0 discourage repetition, while values less than 1.0 encourage it. A value of 1.0 applies no penalty.
 
 Default: `1.1`
+
+### `tts_args` (TTS stages only)
+
+Configuration for Text-to-Speech specific parameters. This section is only applicable to TTS model stages (e.g., `qwen3_tts`).
+
+#### `tts_args.max_instructions_length`
+
+Maximum character length for voice style/emotion instructions. Instructions exceeding this limit will be rejected with a validation error.
+
+Default: `500`
+
+This value can be overridden at runtime using the `--tts-max-instructions-length` CLI parameter when starting the server.

docs/contributing/ci/CI_5levels.md

@@ -0,0 +1,678 @@
+# Multi-Level Automated Testing System Documentation
+
+## Document Overview
+
+This testing system aims to build a complete, efficient, and well-structured quality assurance framework for the development, integration, and release of model services. It draws on the concept of the test pyramid from modern software engineering, progressively expanding testing activities from basic code logic verification to complex end-to-end (E2E) functionality, performance, accuracy, and even long-term stability validation.
+
+Through five levels (L1-L5) and common (Common) specifications, the system clarifies the testing objectives, scope, execution frequency, and required resources for different development stages (e.g., each commit, PR merge, daily build, pre-release). This ensures that models meet high standards for functionality, performance, and reliability across various deployment scenarios (online serving and offline inference).
+
+<table>
+  <thead>
+    <tr>
+      <th>Level</th>
+      <th>Scope & Focus</th>
+      <th>Time Cost</th>
+      <th>Test Dir</th>
+      <th>Doc</th>
+      <th>Frequency</th>
+      <th>Hardware</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td rowspan="2"><strong>Common</strong></td>
+      <td>Contribution Guideline & PR checklist</td>
+      <td>/</td>
+      <td>/</td>
+      <td>.github/PULL_REQUEST_TEMPLATE.md <a href="../tests_style/"> Test Style (PR Checklist)</a></td>
+      <td>/</td>
+      <td>/</td>
+    </tr>
+    <tr>
+      <td>CI Failure Description</td>
+      <td>/</td>
+      <td>/</td>
+      <td><a href="../failures/"> CI Failures</a></td>
+      <td>/</td>
+      <td>/</td>
+    </tr>
+    <tr>
+      <td><strong>L1</strong><br>(Unit & Logic)</td>
+      <td>Unit tests for components like entrypoints, models</td>
+      <td rowspan="2">&lt;15min</td>
+      <td>/tests/{component_name}/test_xxx</td>
+      <td>
+        <a href="#chapter-1-l1-l2-level-testing-unit-testing-and-basic-end-to-end-verification">Chapter 1</a><br>
+        Section 1 L1&amp;L2: Purpose, Test Content, Directory Location, Example
+      </td>
+      <td>PR with ready label (also can run locally)</td>
+      <td>CPU</td>
+    </tr>
+    <tr>
+      <td><strong>L2</strong><br>(E2E across models & GPU-required UT)</td>
+      <td>Online & Offline (basic deployment scenarios):<br>dummy, normal inference function (output format, stream), some instance startup UT</td>
+      <td>
+        /tests/e2e/online_serving/test_{model_name}.py<br>
+        /tests/e2e/offline_inference/test_{model_name}.py
+      </td>
+      <td>
+        <a href="#chapter-1-l1-l2-level-testing-unit-testing-and-basic-end-to-end-verification">Chapter 1</a><br>
+        L1&amp;L2: Purpose, Test Content, Directory Location, Example
+      </td>
+      <td>PR with ready label</td>
+      <td>GPU</td>
+    </tr>
+    <tr>
+      <td><strong>L3</strong><br>(Important Perf & Integration & Accuracy)</td>
+      <td>Online & Offline (multiple deployment scenarios):<br>real model, normal inference function, normal accuracy</td>
+      <td>&lt;30min</td>
+      <td>
+        /tests/e2e/online_serving/test_{model_name}_expansion.py<br>
+        /tests/e2e/offline_inference/test_{model_name}_expansion.py
+      </td>
+      <td>
+        <a href="#chapter-2-l3-level-testing-core-integration-performance-and-accuracy-verification">Chapter 2</a><br>
+        L3: Purpose, Test Content, Directory Location, Example
+      </td>
+      <td>PR Merged (Also run L1&L2 Tests)</td>
+      <td>GPU</td>
+    </tr>
+    <tr>
+      <td><strong>L4</strong><br>(Perf & Integration & Accuracy)</td>
+      <td>Online & Offline: full functional scenarios + performance test + doc test</td>
+      <td>&lt;3 hour</td>
+      <td>
+        <strong>Full Function:</strong><br>
+        /tests/e2e/online_serving/test_{model_name}_expansion.py<br>
+        /tests/e2e/offline_inference/test_{model_name}_expansion.py<br>
+        <strong>Performance:</strong><br>
+        /tests/e2e/perf/nightly.json<br>
+        <strong>Doc Test:</strong><br>
+        tests/example/online_serving/test_{model_name}.py<br>
+        tests/example/offline_inference/test_{model_name}.py
+      </td>
+      <td>
+        <a href="#chapter-3-l4-level-testing-full-functionality-performance-and-documentation-testing">Chapter 3</a><br>
+        L4: Purpose, Test Content, Directory Location, Example
+      </td>
+      <td>Nightly</td>
+      <td>GPU</td>
+    </tr>
+    <tr>
+      <td><strong>L5</strong><br>(Stability & Reliability)</td>
+      <td>Online & Offline: long-term stability test + reliability test</td>
+      <td> Depends on reality </td>
+      <td>
+        <strong>Stability:</strong><br>
+        tests/e2e/stability/weekly.json<br>
+        <strong>Reliability:</strong><br>
+        tests/e2e/reliability/test_{model_name}.py
+      </td>
+      <td>
+        <a href="#chapter-4-l5-level-testing-stability-and-reliability-testing">Chapter 4</a><br>
+        L5: Purpose, Test Content, Directory Location, Example
+      </td>
+      <td>Weekly / Days before Release</td>
+      <td>GPU</td>
+    </tr>
+  </tbody>
+</table>
+
+
+---
+<details>
+<summary> The folder structure for tests file based on the 5 levels design</summary>
+Legend: `✅` = test exists, `⬜` = suggested to add.
+```
+vllm_omni/                                    tests/
+├── config/                             →     ├── config/
+│   ├── model.py                              │   └── test_model.py                    ⬜
+│   └── lora.py                               │   └── test_lora.py                      ⬜
+│
+├── core/                               →     ├── core/
+│   └── sched/                                 │   └── sched/
+│       ├── omni_ar_scheduler.py               │       ├── test_omni_ar_scheduler.py    ⬜
+│       ├── omni_generation_scheduler.py       │       ├── test_omni_generation_scheduler.py  ⬜
+│       └── output.py                          │       └── test_output.py               ✅ currently in entrypoints/test_omni_new_request_data.py (tests output.OmniNewRequestData)
+│
+├── diffusion/                          →     ├── diffusion/
+│   ├── diffusion_engine.py                    │   ├── test_diffusion_engine.py          ⬜
+│   ├── attention/                             │   ├── attention/
+│   │   ├── layer.py                            │   │   ├── test_attention_sp.py         ✅
+│   │   └── backends/                           │   │   └── test_flash_attn.py           ✅
+│   ├── distributed/                           │   ├── distributed/
+│   │   └── ...                                 │   │   ├── test_comm.py                 ✅
+│   │                                            │   │   ├── test_cfg_parallel.py        ✅
+│   │                                            │   │   └── test_sp_plan_hooks.py       ✅
+│   ├── lora/                                   │   ├── lora/
+│   │   └── ...                                 │   │   ├── test_base_linear.py          ✅
+│   │                                            │   │   └── test_lora_manager.py        ✅
+│   ├── models/                                 │   ├── models/
+│   │   ├── qwen_image/                         │   │   ├── qwen_image/                 (e2e coverage)
+│   │   ├── z_image/                            │   │   └── z_image/
+│   │   └── ...                                 │   │       └── test_zimage_tp_constraints.py  ✅
+│   └── worker/                                 │   └── worker/
+│       ├── diffusion_worker.py                 │       └── test_diffusion_worker.py   ✅ file at diffusion/test_diffusion_worker.py
+│       └── diffusion_model_runner.py            │
+│
+├── distributed/                         →     ├── distributed/
+│   └── omni_connectors/                         │   └── omni_connectors/
+│       ├── adapter.py                           │       ├── test_adapter_and_flow.py   ✅
+│       ├── kv_transfer_manager.py               │       ├── test_basic_connectors.py   ✅
+│       ├── connectors/                           │       ├── test_kv_flow.py             ✅
+│       └── utils/                               │       └── test_omni_connector_configs.py  ✅
+│
+├── engine/                             →     ├── engine/
+│   ├── input_processor.py                      │   ├── test_input_processor.py         ⬜  (no processor.py in source)
+│   ├── output_processor.py                     │   └── test_output_processor.py         ⬜
+│   └── arg_utils.py                            │   └── test_arg_utils.py               ⬜
+│
+├── entrypoints/                        →     ├── entrypoints/
+│   ├── omni.py                                 │   ├── test_omni.py                    ⬜  (E2E covered by e2e/offline, e2e/online)
+│   ├── omni_llm.py                             │   ├── test_omni_llm.py                ✅
+│   ├── omni_stage.py                            │   ├── test_omni_stage.py              ⬜  (partial in test_omni_stage_diffusion_config.py)
+│   ├── omni_diffusion.py                       │   ├── test_omni_diffusion.py          ✅
+│   ├── async_omni.py                            │   ├── test_async_omni.py              ✅ actually in e2e/online_serving/test_async_omni.py
+│   ├── async_omni_diffusion.py                 │   ├── test_async_omni_diffusion_config.py  ✅
+│   ├── stage_utils.py                          │   ├── test_stage_utils.py            ✅
+│   ├── cli/                                     │   ├── cli/                           (benchmarks/test_serve_cli.py covers CLI serve)
+│   │   └── ...                                  │   │   └── test_*.py                  ⬜
+│   └── openai/                                  │   └── openai_api/                    # maps to entrypoints/openai/
+│       ├── api_server.py                        │       ├── test_api_server.py         ⬜  (e2e indirect coverage)
+│       ├── serving_chat.py                       │       ├── test_serving_chat_sampling_params.py  ✅
+│       ├── serving_speech.py                     │       ├── test_serving_speech.py     ✅
+│       └── image_api_utils.py                   │       └── test_image_server.py      ✅
+│
+├── inputs/                             →     ├── inputs/
+│   ├── data.py                                 │   ├── test_data.py                   ⬜
+│   ├── parse.py                                │   ├── test_parse.py                 ⬜
+│   └── preprocess.py                            │   └── test_preprocess.py            ✅ currently in entrypoints/test_omni_input_preprocessor.py
+│
+├── model_executor/                     →     ├── model_executor/
+│   ├── layers/                                  │   ├── layers/
+│   │   └── mrope.py                             │   │   └── test_mrope.py              ⬜
+│   ├── model_loader/                            │   ├── model_loader/
+│   │   └── weight_utils.py                      │   │   └── test_weight_utils.py      ⬜
+│   ├── models/                                  │   ├── models/
+│   │   ├── qwen2_5_omni/                         │   │   ├── qwen2_5_omni/
+│   │   │   ├── qwen2_5_omni_thinker.py           │   │   │   ├── test_audio_length.py  ✅
+│   │   │   ├── qwen2_5_omni_talker.py            │   │   │   ├── test_qwen2_5_omni_thinker.py  ⬜
+│   │   │   └── qwen2_5_omni_token2wav.py         │   │   │   ├── test_qwen2_5_omni_talker.py  ⬜
+│   │   └── qwen3_omni/                          │   │   │   └── test_qwen2_5_omni_token2wav.py  ⬜
+│   │       └── ...                               │   │   └── qwen3_omni/
+│   ├── stage_configs/                           │   │       └── test_*.py              ⬜
+│   │   └── *.yaml                               │   └── stage_configs/                 (used by e2e, test_*.py can be added)  ⬜
+│   └── stage_input_processors/                  │   └── stage_input_processors/
+│       └── ...                                  │       └── test_*.py                 ⬜
+│
+├── sample/                             →     ├── sample/
+│   └── __init__.py                             │   └── test_*.py                      ⬜
+│
+├── utils/                              →     ├── utils/
+│   └── __init__.py                             │   └── test_*.py                       ⬜  (no platform_utils.py currently)
+│
+├── worker/                             →     ├── worker/
+│   ├── gpu_ar_model_runner.py                  │   ├── test_gpu_ar_model_runner.py    ⬜
+│   ├── gpu_ar_worker.py                        │   ├── test_gpu_ar_worker.py           ⬜
+│   ├── gpu_generation_model_runner.py          │   ├── test_gpu_generation_model_runner.py  ✅
+│   ├── gpu_generation_worker.py                │   ├── test_gpu_generation_worker.py  ⬜
+│   ├── gpu_model_runner.py                     │   ├── test_omni_gpu_model_runner.py   ✅
+│   └── mixins.py                               │   └── (npu under platforms/npu/worker/)  # not worker/npu/
+│
+├── platforms/                          →     (no tests/platforms/, e2e and stage_configs provide indirect coverage)
+│   ├── cuda/
+│   ├── npu/worker/                             # NPU worker here, not vllm_omni/worker/npu/
+│   ├── rocm/
+│   └── xpu/worker/
+│
+├── outputs.py                          →     test_outputs.py                         ✅ (at tests root)
+├── (logger, patch, request, version)    →     (no corresponding unit test)
+│
+└── e2e (tests side only)               →     ├── e2e/
+                                               ├── online_serving/                     ✅ non-empty
+                                               │   ├── test_async_omni.py
+                                               │   ├── test_qwen3_omni.py
+                                               │   ├── test_qwen3_omni_expansion.py
+                                               │   ├── test_image_gen_edit.py
+                                               │   ├── test_images_generations_lora.py
+                                               │   └── stage_configs/
+                                               └── offline_inference/                  ✅
+                                                   ├── test_qwen2_5_omni.py
+                                                   ├── test_qwen3_omni.py
+                                                   ├── test_bagel_text2img.py
+                                                   ├── test_t2i_model.py
+                                                   ├── test_t2v_model.py
+                                                   ├── test_ovis_image.py
+                                                   ├── test_zimage_tensor_parallel.py
+                                                   ├── test_cache_dit.py
+                                                   ├── test_teacache.py
+                                                   ├── test_stable_audio_model.py
+                                                   ├── test_diffusion_cpu_offload.py
+                                                   ├── test_diffusion_layerwise_offload.py
+                                                   ├── test_diffusion_lora.py
+                                                   ├── test_sequence_parallel.py
+                                                   └── stage_configs/
+                                                       ├── qwen2_5_omni_ci.yaml
+                                                       ├── qwen3_omni_ci.yaml
+                                                       ├── bagel_*.yaml
+                                                       └── npu/, rocm/, etc.
+```
+
+
+</details>
+
+
+## Common Specifications
+
+Before entering specific testing levels, the project establishes two common specifications aimed at standardizing the development process and quickly locating issues.
+
+1.  ***PR Checklist ([Tests Style](../ci/tests_style.md))***: This template defines the self-check items that must be completed before submitting a code review (Pull Request). It ensures that each code change meets basic requirements such as code style, dependency updates, and documentation synchronization before entering the automated testing pipeline, serving as the first manual line of defense for quality assurance.
+2.  ***CI Failure Explanation ([CI Failures](../ci/failures.md))***: This document archives and explains common failure patterns in the Continuous Integration (CI) pipeline, error log interpretation, and preliminary troubleshooting steps. It helps developers and testers quickly diagnose the causes of automated test failures, improving problem-solving efficiency.
+
+## Chapter 1: L1 & L2 Level Testing - Unit Testing and Basic End-to-End Verification
+
+### 1.1 Testing Purpose
+
+L1 and L2 level testing form the foundation of the quality assurance system. L1 level testing focuses on verifying the internal logic correctness of code units (e.g., functions, classes), ensuring each independent component behaves as designed.
+
+L2 level testing builds upon L1 by introducing GPU resources and verifying that the end-to-end (E2E) process of the model in basic deployment scenarios is smooth. For example, it uses dummy models to confirm that core interfaces like the inference pipeline, output format, and streaming response work properly. The common goal of these two levels is to provide developers with rapid feedback, discovering and fixing issues early in the development cycle  .
+
+
+
+### 1.2 Testing Content and Scope
+
+-   ***L1 (Unit & Logic Testing)***:
+-   -   ***Scope***: Tests internal functions and methods of core components such as `entrypoints`, `models`.
+    -   ***Focus***: Branch coverage, exception handling, algorithm logic correctness. Does not involve external dependencies or the complete service stack.
+    -   ***Time Cost***: Execution time is controlled within ***15 minutes*** to ensure fast feedback.
+-   ***L2 (Basic End-to-End Testing)***:
+-   -   ***Scope***: Covers two basic deployment scenarios: `online` (serving) and `offline` (inference).
+    -   ***Focus***: Uses `dummy` models or lightweight real models to verify that the entire chain from request input to result output works normally, including output data structure, streaming (stream) support, etc. Also includes some unit tests that require launching independent service instances.
+    -   ***Characteristic***: Requires ***GPU*** resources to perform model computations.
+
+### 1.3 Test Directory and Execution Files
+
+A clear directory structure is key to managing test cases efficiently.
+
+-   ***L1 Test Directory***: `/tests/{component_name}/test_xxx.py`
+-   -   Here, `{component_name}` corresponds to modules in the source code, such as `distributed`, `entrypoints`, etc., and `test_xxx.py` is the specific test file.
+-   ***L2 Test Directory***:
+-   -   Online Serving: `/tests/e2e/online_serving/test_{model_name}.py`
+    -   Offline Inference: `/tests/e2e/offline_inference/test_{model_name}.py`
+
+### 1.4 Execution Method and Example
+
+-   ***Trigger Timing***: **`PR with ready label`**. That is, when a developer adds a "ready for review" or similar label to a PR on platforms like GitHub, L1 and L2 tests are automatically triggered.
+-   ***Execution Environment***: L1 uses ***CPU*** environment; L2 requires ***GPU*** environment.
+-   ***Script Example***:
+
+<details>
+<summary> L1 Test Examples</summary>
+
+Examples from `tests/model_executor/models/qwen2_5_omni/test_audio_length.py`
+```python
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+def test_resolve_max_mel_frames_default():
+    from vllm_omni.model_executor.models.qwen2_5_omni.audio_length import resolve_max_mel_frames
+
+    assert resolve_max_mel_frames(None, default=30000) == 30000
+    assert resolve_max_mel_frames(None, default=6000) == 6000
+
+
+def test_resolve_max_mel_frames_explicit():
+    from vllm_omni.model_executor.models.qwen2_5_omni.audio_length import resolve_max_mel_frames
+
+    # Explicit argument always wins over default
+    assert resolve_max_mel_frames(123, default=30000) == 123
+    assert resolve_max_mel_frames(6000, default=30000) == 6000
+    assert resolve_max_mel_frames(0, default=30000) == 0
+
+
+@pytest.mark.parametrize("repeats", [2, 4])
+@pytest.mark.parametrize("code_len", [0, 1, 32768])
+@pytest.mark.parametrize("max_mel_frames", [None, -1, 0, 1, 6000, 30000])
+def test_cap_and_align_mel_length_no_mismatch(repeats, code_len, max_mel_frames):
+    """Guard that any max_mel_frames yields a mel length aligned to repeats, and
+    consistent with the truncated code length (prevents concat mismatch).
+    """
+    from vllm_omni.model_executor.models.qwen2_5_omni.audio_length import cap_and_align_mel_length
+
+    target_code_len, target_mel_len = cap_and_align_mel_length(
+        code_len=code_len,
+        repeats=repeats,
+        max_mel_frames=max_mel_frames,
+    )
+
+    assert isinstance(target_code_len, int)
+    assert isinstance(target_mel_len, int)
+
+    if code_len == 0:
+        assert target_code_len == 0
+        assert target_mel_len == 0
+        return
+
+    assert target_code_len >= 1
+    assert target_mel_len >= repeats
+    assert target_mel_len % repeats == 0
+    assert target_mel_len == target_code_len * repeats
+    assert target_code_len <= code_len
+
+    if max_mel_frames is not None and int(max_mel_frames) > 0 and int(max_mel_frames) >= repeats:
+        assert target_mel_len <= int(max_mel_frames)
+```
+</details>
+
+<details>
+<summary> L2 Test Examples</summary>
+You can refer to Test Examples in Chapter 2 to see example test cases that incorporate both L2 and L3 testing logic.
+</details>
+
+-   -   ***Run Command***:
+
+    `pytest -s -v /tests/e2e/online_serving/test_{model_name}.py`
+    `pytest -s -v -m 'core_model and cpu' --run-level=core_model`
+
+## Chapter 2: L3 Level Testing - Core Integration, Performance, and Accuracy Verification
+
+### 2.1 Testing Purpose
+
+L3 level testing executes after code is merged into the main branch. Its core purpose is to verify the integration effect, key performance indicators, and output accuracy of ***real models*** in ***multiple deployment scenarios***
+
+. It acts as the "quality gatekeeper" for the main branch, ensuring that no merge breaks the core capabilities of the model service. Testing needs to provide clear conclusions within a relatively short time (<30min), balancing test depth with feedback speed.
+
+
+
+### 2.2 Testing Content and Scope
+
+-   ***Deployment Scenarios***: Covers richer `online` and `offline` deployment configurations, which may include different hardware configurations, batch sizes, concurrency levels, etc.
+-   ***Core Verification***:
+-   1.  ***Inference Functionality***: Ensures real models can perform forward computation normally and return results.
+    2.  ***Accuracy Compliance***: Verifies that the model's evaluation metrics (e.g., accuracy) meet the expected baseline, preventing code changes from introducing accuracy issues.
+    3.  ***Important Performance***: Verifies whether performance (e.g., P99 latency, throughput) in core scenarios meets preset thresholds.
+
+### 2.3 Test Directory and Execution Files
+
+-   ***Functional Testing***:
+-   -   Online Serving: `/tests/e2e/online_serving/test_{model_name}_expansion.py`
+    -   Offline Inference: `/tests/e2e/offline_inference/test_{model_name}_expansion.py`
+    -   (Note: `_expansion.py` likely means it contains more comprehensive scenario cases compared to L2 tests).
+
+### 2.4 Execution Method and Example
+
+-   ***Trigger Timing***: **`PR Merged`**. Automatically triggered after code review is approved and merged into the main branch.
+-   ***Execution Environment***: ***GPU*** servers.
+-   ***Script Example***:
+
+???+ example "Test Examples"
+
+    **2.4.1 Mark Declaration Section**
+
+    ```python
+    @pytest.mark.advanced_model
+    @pytest.mark.core_model
+    @pytest.mark.parametrize("omni_server", test_params, indirect=True)
+    ```
+
+    **Explanation**:
+
+    @pytest.mark.advanced_model: Marks the test as L3 or L4 level, indicating that this test case performs deep validation, using real models for performance, integration, and accuracy testing. This forms a "basic-advanced" correspondence with the core_model mark at the L2 level.
+
+    @pytest.mark.core_model: Marks the test as L1 or L2 level, indicating that this test case validates the basic functionality of the core model. It uses mock weights and only checks if the relevant interface functions correctly.
+
+    @pytest.mark.parametrize: A parameterization decorator that allows abstracting test data into parameters, enabling reuse of the same test logic across different data configurations. indirect=True indicates that parameters will be passed to the fixture for processing.
+
+    **Notes**: If you believe the test case only needs to execute basic run logic at the PR-level CI, you can mark it only with @pytest.mark.core_model. If you believe it only needs to execute deep validation run logic at the merge or nightly level, you can mark it only with @pytest.mark.advanced_model. If you believe the test case needs to accommodate both basic run and deep validation test logic, you should mark it with both @pytest.mark.core_model and @pytest.mark.advanced_model.
+
+    **2.4.2 Test Function Definition and Documentation**
+
+    ```python
+    def test_mix_to_text_audio_001(omni_server, openai_client) -> None:
+        """
+        Test multi-modal input processing and text/audio output generation via OpenAI API.
+        Deploy Setting: default yaml
+        Input Modal: text + audio + video + image
+        Output Modal: text + audio
+        Input Setting: stream=True
+        Datasets: single request
+        """
+    ```
+
+    **Explanation**:
+
+    **Function Naming Convention**: Uses the test_ prefix, describes the test scenario mix_to_text_audio, and the number 001 indicates the first test case for this scenario.
+
+    **Parameter Explanation**:
+
+    omni_server: Omni server instance obtained via fixture, containing model information and configuration.
+
+    openai_client: Unified OpenAI client processing instance, encapsulating request sending and response validation logic.
+
+    Docstring: Describes the test purpose, deployment settings, input/output modalities, streaming settings, and dataset type in detail, providing clear context for test maintenance.
+
+    **2.4.3 Multimodal Data Preparation**
+
+    ```python
+    video_data_url = f"data:video/mp4;base64,{generate_synthetic_video(224, 224, 300)['base64']}"
+    image_data_url = f"data:image/jpeg;base64,{generate_synthetic_image(224, 224)['base64']}"
+    audio_data_url = f"data:audio/wav;base64,{generate_synthetic_audio(5, 1)['base64']}"
+    ```
+
+    **Explanation**:
+
+    **Data Generation Functions**: Use the generate_synthetic_* series of functions to generate synthetic test data, avoiding reliance on external resources and ensuring test reproducibility and stability.
+
+    **Parameter Explanation**:
+
+    Video: width, height, duration_frames
+
+    Image: width, height
+
+    Audio: duration_seconds, channels
+
+    **2.4.4 Request Configuration and Keyword Validation**
+
+    ```python
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "key_words": {
+            "audio": ["water", "cricket"],
+            "video": ["sphere", "globe", "circle", "round"],
+            "image": ["square", "quadrate"],
+            "text": ["beijing"]
+        },
+    }
+    ```
+
+    **Explanation**:
+
+    **Model Specification**: Uses omni_server.model to ensure the test aligns with the model configured on the server.
+
+    **Keyword Validation Mechanism**: This is an innovative design of the template to address the specific needs of multimodal testing:
+
+    Audio Keywords: Validate whether the generated text's description of audio content contains expected elements (e.g., "water" for water sounds, "cricket" for cricket sounds). If you provide multiple keywords, the validation is considered successful if at least one keyword is present.
+
+    **Video Keywords**: Validate whether the generated text's description of video content contains expected elements. If you provide multiple keywords, the validation is considered successful if at least one keyword is present.
+
+    Image Keywords: Validate whether the generated text's description of image content contains expected elements. If you provide multiple keywords, the validation is considered successful if at least one keyword is present.
+
+    Text Keywords: Validate whether the generated text contains expected elements. If you provide multiple keywords, the validation is considered successful if at least one keyword is present.
+
+    **2.4.5 Request Execution**
+
+    ```python
+    openai_client.send_request(request_config, request_num=1)
+    ```
+
+    **Explanation**:
+
+    **Unified Client**: Uses the OpenAIClientHandler instance to send requests. This client encapsulates error handling, retry mechanisms, and response validation logic.
+
+    **Single Request**: The comment clearly states this is a single-request completion test. For concurrent testing, it can be extended to multiple requests using request_num = n.
+
+    **Implicit Validation**: The send_request method internally includes validation logic dynamically selected based on the --run-level parameter: core_model performs basic validation, while advanced_model performs deep validation.
+
+-   ***Run Command***: `pytest -s -v /tests/e2e/online_serving/test_{model_name}.py -m advanced_model --run-level=advanced_model`
+
+## Chapter 3: L4 Level Testing - Full Functionality, Performance, and Documentation Testing
+
+### 3.1 Testing Purpose
+
+L4 level testing is a comprehensive quality audit before a version release. It expands upon L3, executing ***full*** functional scenarios, conducting systematic ***performance stress tests***, and simultaneously verifying the correctness of accompanying ***example documentation***. Its purpose is to perform deep validation of the system during off-peak nighttime hours, providing quality trend reports for daytime development and data support for release decisions.
+
+
+
+### 3.2 Testing Content and Scope
+
+-   ***Full Functionality Testing***: Executes all test cases defined in `test_{model_name}_expansion.py`, covering all implemented features, positive flows, boundary conditions, and exception handling.
+-   ***Performance Testing***: Uses the `/tests/e2e/perf/nightly.json` configuration file to drive performance testing tools for stress, load, and endurance tests, collecting metrics like throughput, response time, and resource utilization.
+-   ***Documentation Testing***: Verifies whether the example code provided to users is runnable and its results match the description.
+
+### 3.3 Test Directory and Execution Files
+
+-   ***Functional Testing***: Same directories as L3.
+-   ***Performance Test Configuration***: `/tests/e2e/perf/nightly.json`
+-   ***Documentation Example Tests***:
+-   -   `tests/example/online_serving/test_{model_name}.py`
+    -   `tests/example/offline_inference/test_{model_name}.py`
+
+### 3.4 Execution Method and Example
+
+-   ***Trigger Timing***: **`Nightly`**, automatically executed every night.
+-   ***Execution Environment***: ***GPU*** server clusters to meet the resource demands of performance testing.
+-   ***Script Example***:
+<details>
+<summary> Test Examples</summary>
+When you want to add L4-level performance test cases, you can refer to the following format for case addition in tests/perf/tests/test.json:
+
+```JSON
+{
+    "test_name": "test_qwen3_omni",
+    "server_params": {
+        "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
+        "stage_config_name": "qwen3_omni.yaml"
+    },
+    "benchmark_params": [
+        {
+            "dataset_name": "random",
+            "num_prompts": [10, 20],
+            "request_rate": [0.5, 1],
+            "random_input_len": 2500,
+            "random_output_len": 900,
+            "ignore_eos": true,
+            "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
+            "baseline": {
+                "mean_ttft_ms": 100000,
+                "mean_audio_ttfp_ms": 100000,
+                "mean_audio_rtf": 100000
+            }
+        }
+    ]
+}
+```
+
+#### Parameter Explanation
+
+***Overview***
+
+| Field            | Required | Description                                                     |
+| ---------------- | -------- | --------------------------------------------------------------- |
+| test_name        | Yes      | Unique identifier for the test case                             |
+| server_params    | Yes      | Server-side configuration parameters                            |
+| benchmark_params | Yes      | Benchmark running parameters (supports multiple configurations) |
+
+#### server_params Configuration
+
+##### Basic Parameters
+
+| Parameter         | Required | Example                            | Description                   |
+| ----------------- | -------- | ---------------------------------- | ----------------------------- |
+| model             | Yes      | "Qwen/Qwen3-Omni-30B-A3B-Instruct" | Model name or path            |
+| stage_config_name | Yes      | "qwen3_omni.yaml"                  | Stage configuration file name |
+
+##### Dynamic Configuration (update/delete)
+
+Supports incremental modifications based on the basic configuration:
+
+| Operation | Description                          |
+| --------- | ------------------------------------ |
+| update    | Update or add configuration items    |
+| delete    | Delete specified configuration items |
+
+***Example***:
+```
+"update": {
+    "async_chunk": true,  // Enable asynchronous chunk processing
+    "stage_args": {
+        "0": {
+            "engine_args.custom_process_next_stage_input_func": "vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_async_chunk"
+        }
+    }
+},
+"delete": {
+    "stage_args": {
+        "2": ["custom_process_input_func"]  // Delete this configuration for stage 2
+    }
+}
+```
+#### benchmark_params Configuration
+
+You can add any benchmark running parameters you need here. For all optional parameters, refer to the [benchmark documentation](https://github.com/vllm-project/vllm-omni/blob/main/docs/cli/bench/serve.md). General modifications are as follows:
+
+1.  Change the ---xxx-xx-xx running parameters to xxx_xx_xx format and fill them as keys in the JSON file.
+2.  For boolean variables in the running parameters, modify them to forms such as ignore_eos: true/false and fill them into the JSON file.
+3.  Add the baseline parameter to specify the required validation values, ensuring the validation metric names match those in the result.json generated by the benchmark.
+4.  The qps and concurrency modes are mutually exclusive. For detailed explanations, see the table below:
+
+| Parameter       | Type        | Required | Example/Values  | Description                                                                                                                                                                                                                                                          |
+| --------------- | ----------- | -------- | --------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| num_prompts     | int / array | Yes      | 10,[10, 20, 30] | Number of requests. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of qps or max_concurrency, e.g., [10,10,10]. If an array is used, its length must match the number of qps or max_concurrency. |
+| request_rate    | int / array | No       | 1, [1, 2, 3]    | Queries per second. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of num_prompts, e.g., [1,1,1]. If an array is used, its length must match the number of num_prompts.                          |
+| max_concurrency | int / array | No       | 1, [1, 2, 3]    | Queries per second. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of num_prompts, e.g., [1,1,1]. If an array is used, its length must match the number of num_prompts.                          |
+</details>
+
+-   -   ***Run Command***: (Specific commands would depend on the performance testing tool and configuration defined in `nightly.json`).
+
+## Chapter 4: L5 Level Testing - Stability and Reliability Testing
+
+### 4.1 Testing Purpose
+
+L5 level testing focuses on the performance of model services under ***long-running*** and ***abnormal fault*** scenarios. It aims to uncover deep-seated issues that only manifest under sustained pressure or extreme conditions, such as memory leaks, resource contention, gradual performance degradation, and lack of fault tolerance mechanisms. This is the final, yet crucial, line of defense for ensuring service high availability and production environment robustness.
+
+
+
+### 4.2 Testing Content and Scope
+
+-   ***Long-term Stability (Stability) Testing***: Uses the `tests/e2e/stability/weekly.json` configuration to run the service under moderate load for an extended period (e.g., over 12 hours), monitoring whether metrics like memory/VRAM usage, response time, and throughput degrade over time, and whether the service process remains stable.
+-   ***Reliability Testing***: Uses `tests/e2e/reliability/test_{model_name}.py` to actively simulate various fault and abnormal scenarios, such as: dependent service interruption, abnormal input data, network flicker, hardware resource preemption, etc., to verify the system's fault tolerance, self-healing, and graceful degradation capabilities.
+
+### 4.3 Test Directory and Execution Files
+
+-   ***Stability Test Configuration***: `tests/e2e/stability/weekly.json`
+-   ***Reliability Test Suite***: `tests/e2e/reliability/test_{model_name}.py`
+
+### 4.4 Execution Method and Example
+
+-   ***Trigger Timing***: **`Weekly`** (weekly) or **`Days before Release`** (several days before a major release). Due to long execution times, the frequency is lower.
+-   ***Execution Environment***: ***GPU*** servers, requiring a stable and exclusive testing environment.
+-   ***Script Example***:
+<details>
+<summary> Test Examples</summary>
+```python
+# WIP
+```
+</details>
+
+-   -   ***Stability***: (Execution would be driven by the configuration in `weekly.json`).
+    -   ***Reliability***: `pytest -s -v tests/e2e/reliability/test_{model_name}.py`
+
+## Summary
+
+This multi-level testing system achieves continuous, progressive validation of model service quality by tightly integrating testing activities with the development workflow (commit, review, merge, release). From rapid unit testing to comprehensive end-to-end testing, and further to in-depth performance, stability, and reliability verification, each level has clear objectives, collectively building a robust quality protection net. By following this system, teams can deliver high-quality, highly reliable model services more efficiently.

docs/contributing/ci/failures.md

@@ -1,4 +1,10 @@
 # CI Failures
 
-What should I do when a CI job fails on my PR, but I don't think my PR caused
-the failure?
+What should I do when a CI job fails on my PR, but I don't think my PR caused the failure?
+
+
+## Common Case 1:
+
+🚨 Error: The command was interrupted by a signal: signal: terminated
+
+Reason: The test is terminated due to exceed the time limits. Sometimes rebuild the test will pass.

docs/contributing/ci/tests_markers.md

@@ -5,33 +5,34 @@ By adding markers before test functions, tests can later be executed uniformly b
 ## Current Markers
 Defined in `pyproject.toml`:
 
-| Marker             | Description                                             |
-| ------------------ | ------------------------------------------------------- |
-| `core_model`       | Core model tests (run in each PR)                       |
-| `diffusion`        | Diffusion model tests                                   |
-| `omni`             | Omni model tests                                        |
-| `cache`            | Cache backend tests                                     |
-| `parallel`         | Parallelism/distributed tests                           |
-| `cpu`              | Tests that run on CPU                                   |
-| `gpu`              | Tests that run on GPU (auto-added)                      |
-| `cuda`             | Tests that run on CUDA (auto-added)                     |
-| `rocm`             | Tests that run on AMD/ROCm (auto-added)                 |
-| `npu`              | Tests that run on NPU/Ascend (auto-added)               |
-| `H100`             | Tests that require H100 GPU                             |
-| `L4`               | Tests that require L4 GPU                               |
-| `MI325`            | Tests that require MI325 GPU (AMD/ROCm)                 |
-| `A2`               | Tests that require A2 NPU                               |
-| `A3`               | Tests that require A3 NPU                               |
-| `distributed_cuda` | Tests that require multi cards on CUDA platform         |
-| `distributed_rocm` | Tests that require multi cards on ROCm platform         |
-| `distributed_npu`  | Tests that require multi cards on NPU platform          |
-| `skipif_cuda`      | Skip if the num of CUDA cards is less than the required |
-| `skipif_rocm`      | Skip if the num of ROCm cards is less than the required |
-| `skipif_npu`       | Skip if the num of NPU cards is less than the required  |
-| `slow`             | Slow tests (may skip in quick CI)                       |
-| `benchmark`        | Benchmark tests                                         |
+| Marker             | Description                                               |
+| ------------------ | --------------------------------------------------------- |
+| `core_model`       | L1&L2 tests (run in each PR)                              |
+| `advanced_model`   | L3&L4 level tests (run in each merge or nightly)          |
+| `diffusion`        | Diffusion model tests                                     |
+| `omni`             | Omni model tests                                          |
+| `cache`            | Cache backend tests                                       |
+| `parallel`         | Parallelism/distributed tests                             |
+| `cpu`              | Tests that run on CPU                                     |
+| `gpu`              | Tests that run on GPU *                                   |
+| `cuda`             | Tests that run on CUDA *                                  |
+| `rocm`             | Tests that run on AMD/ROCm *                              |
+| `npu`              | Tests that run on NPU/Ascend *                            |
+| `H100`             | Tests that require H100 GPU  *                            |
+| `L4`               | Tests that require L4 GPU *                               |
+| `MI325`            | Tests that require MI325 GPU (AMD/ROCm) *                 |
+| `A2`               | Tests that require A2 NPU *                               |
+| `A3`               | Tests that require A3 NPU *                               |
+| `distributed_cuda` | Tests that require multi cards on CUDA platform *         |
+| `distributed_rocm` | Tests that require multi cards on ROCm platform  *        |
+| `distributed_npu`  | Tests that require multi cards on NPU platform  *         |
+| `skipif_cuda`      | Skip if the num of CUDA cards is less than the required * |
+| `skipif_rocm`      | Skip if the num of ROCm cards is less than the required * |
+| `skipif_npu`       | Skip if the num of NPU cards is less than the required *  |
+| `slow`             | Slow tests (may skip in quick CI)                         |
+| `benchmark`        | Benchmark tests                                           |
 
-For those markers shown as auto-added, they will be added by the `@hardware_test` decorator.
+\* Means those markers are auto-added, and they will be added by the `@hardware_test` decorator.
 
 ### Example usage for markers
 
@@ -71,10 +72,7 @@ This decorator is intended to make hardware-aware, cross-platform test authoring
    Support for `skipif_rocm` and `skipif_npu` will be implemented later.
 
 
-5. **Runs each test in a new process**  
-   Automatically wraps the distributed test with a decorator (`@create_new_process_for_each_test`) to ensure isolation and compatibility with multi-process hardware backends.
-
-6. **Works with pytest filtering**  
+5. **Works with pytest filtering**  
    Allows tests to be filtered and selected at runtime using standard pytest marker expressions (e.g., `-m "distributed_cuda and L4"`).
 
 #### Example usage for decorator
@@ -94,7 +92,6 @@ This decorator is intended to make hardware-aware, cross-platform test authoring
     ```
 - `res` must be a dict; supported resources: CUDA (L4/H100), ROCm (MI325), NPU (A2/A3)
 - `num_cards` can be int (all platforms) or dict (per platform); defaults to 1 when missing
-- `hardware_test` automatically applies `@create_new_process_for_each_test` for distributed tests.
 - Distributed markers (`distributed_cuda`, `distributed_rocm`, `distributed_npu`) are auto-added for multi-card cases
 - Filtering examples:
     - CUDA only: `pytest -m "distributed_cuda and L4"`

docs/contributing/ci/tests_style.md

@@ -2,8 +2,18 @@
 
 To ensure project maintainability and sustainable development, we encourage contributors to submit test code (unit tests, system tests, or end-to-end tests) alongside their code changes. This document outlines the guidelines for organizing and naming test files.
 
+## Checklist before submitting your test files
+
+1. The file is saved in an appropriate place and the file name is clear.
+2. The coding style follows the requirements outlined below.
+3. All test functions have appropriate pytest markers.
+4. For tests that need run in CI, please ensure it labeled as ``@pytest.mark.core_model` the test is configured under the `./buildkite/` folder.
+
+
 ## Test Types
 
+For more details about our [Five Levels Tests design](../ci/CI_5levels.md).
+
 ### Unit Tests and System Tests
 For unit tests and system tests, we strongly recommend placing test files in the same directory structure as the source code being tested, using the naming convention `test_*.py`.
 
@@ -18,95 +28,141 @@ End-to-end tests verify the complete functionality of a system or component. For
 
 ## Test Directory Structure
 
-The ideal directory structure mirrors the source code organization:
+The ideal directory structure mirrors the source code organization. Legend: `✅` = test exists, `⬜` = suggested to add.
 
 ```
-vllm_omni/                          tests/
-├── config/                    →    ├── config/
-│   └── model.py                    │   └── test_model.py
+vllm_omni/                                    tests/
+├── config/                             →     ├── config/
+│   ├── model.py                              │   └── test_model.py                    ⬜
+│   └── lora.py                               │   └── test_lora.py                      ⬜
 │
-├── core/                      →    ├── core/
-│   └── sched/                      │   └── sched/                    # Maps to core/sched/
-│       ├── omni_ar_scheduler.py    │       ├── test_omni_ar_scheduler.py
-│       ├── omni_generation_scheduler.py │  ├── test_omni_generation_scheduler.py
-│       └── output.py               │       └── test_output.py
+├── core/                               →     ├── core/
+│   └── sched/                                 │   └── sched/
+│       ├── omni_ar_scheduler.py               │       ├── test_omni_ar_scheduler.py    ⬜
+│       ├── omni_generation_scheduler.py       │       ├── test_omni_generation_scheduler.py  ⬜
+│       └── output.py                          │       └── test_output.py               ✅ currently in entrypoints/test_omni_new_request_data.py (tests output.OmniNewRequestData)
 │
-├── diffusion/                 →    ├── diffusion/
-│   ├── diffusion_engine.py         │   ├── test_diffusion_engine.py
-│   ├── omni_diffusion.py           │   ├── test_omni_diffusion.py
-│   ├── attention/                  │   ├── attention/                # Maps to diffusion/attention/
-│   │   └── backends/               │   │   └── test_*.py
-│   ├── models/                     │   ├── models/                   # Maps to diffusion/models/
-│   │   ├── qwen_image/             │   │   ├── qwen_image/
-│   │   │   └── ...                 │   │   │   └── test_*.py
-│   │   └── z_image/                │   │   └── z_image/
-│   │       └── ...                 │   │       └── test_*.py
-│   └── worker/                     │   └── worker/                   # Maps to diffusion/worker/
-│       └── ...                     │       └── test_*.py
+├── diffusion/                          →     ├── diffusion/
+│   ├── diffusion_engine.py                    │   ├── test_diffusion_engine.py          ⬜
+│   ├── attention/                             │   ├── attention/
+│   │   ├── layer.py                            │   │   ├── test_attention_sp.py         ✅
+│   │   └── backends/                           │   │   └── test_flash_attn.py           ✅
+│   ├── distributed/                           │   ├── distributed/
+│   │   └── ...                                 │   │   ├── test_comm.py                 ✅
+│   │                                            │   │   ├── test_cfg_parallel.py        ✅
+│   │                                            │   │   └── test_sp_plan_hooks.py       ✅
+│   ├── lora/                                   │   ├── lora/
+│   │   └── ...                                 │   │   ├── test_base_linear.py          ✅
+│   │                                            │   │   └── test_lora_manager.py        ✅
+│   ├── models/                                 │   ├── models/
+│   │   ├── qwen_image/                         │   │   ├── qwen_image/                 (e2e coverage)
+│   │   ├── z_image/                            │   │   └── z_image/
+│   │   └── ...                                 │   │       └── test_zimage_tp_constraints.py  ✅
+│   └── worker/                                 │   └── worker/
+│       ├── diffusion_worker.py                 │       └── test_diffusion_worker.py   ✅ file at diffusion/test_diffusion_worker.py
+│       └── diffusion_model_runner.py            │
 │
-├── distributed/               →    ├── distributed/
-│   └── ...                         │   └── test_*.py
+├── distributed/                         →     ├── distributed/
+│   └── omni_connectors/                         │   └── omni_connectors/
+│       ├── adapter.py                           │       ├── test_adapter_and_flow.py   ✅
+│       ├── kv_transfer_manager.py               │       ├── test_basic_connectors.py   ✅
+│       ├── connectors/                           │       ├── test_kv_flow.py             ✅
+│       └── utils/                               │       └── test_omni_connector_configs.py  ✅
 │
-├── engine/                    →    ├── engine/
-│   ├── processor.py                │   ├── test_processor.py
-│   └── output_processor.py         │   └── test_output_processor.py
+├── engine/                             →     ├── engine/
+│   ├── input_processor.py                      │   ├── test_input_processor.py         ⬜  (no processor.py in source)
+│   ├── output_processor.py                     │   └── test_output_processor.py         ⬜
+│   └── arg_utils.py                            │   └── test_arg_utils.py               ⬜
 │
-├── entrypoints/               →    ├── entrypoints/
-│   ├── omni_llm.py                 │   ├── test_omni_llm.py          # UT: OmniLLM core logic (mocked)
-│   ├── omni_stage.py               │   ├── test_omni_stage.py         # UT: OmniStage logic
-│   ├── omni.py                     │   ├── test_omni.py               # E2E: Omni class (offline inference)
-│   ├── async_omni.py               │   ├── test_async_omni.py         # E2E: AsyncOmni class
-│   ├── cli/                        │   ├── cli/                       # Maps to entrypoints/cli/
-│   │   └── ...                     │   │   └── test_*.py
-│   └── openai/                     │   └── openai/                     # Maps to entrypoints/openai/
-│       ├── api_server.py           │       ├── test_api_server.py     # E2E: API server (online serving)
-│       └── serving_chat.py         │       └── test_serving_chat.py
+├── entrypoints/                        →     ├── entrypoints/
+│   ├── omni.py                                 │   ├── test_omni.py                    ⬜  (E2E covered by e2e/offline, e2e/online)
+│   ├── omni_llm.py                             │   ├── test_omni_llm.py                ✅
+│   ├── omni_stage.py                            │   ├── test_omni_stage.py              ⬜  (partial in test_omni_stage_diffusion_config.py)
+│   ├── omni_diffusion.py                       │   ├── test_omni_diffusion.py          ✅
+│   ├── async_omni.py                            │   ├── test_async_omni.py              ✅ actually in e2e/online_serving/test_async_omni.py
+│   ├── async_omni_diffusion.py                 │   ├── test_async_omni_diffusion_config.py  ✅
+│   ├── stage_utils.py                          │   ├── test_stage_utils.py            ✅
+│   ├── cli/                                     │   ├── cli/                           (benchmarks/test_serve_cli.py covers CLI serve)
+│   │   └── ...                                  │   │   └── test_*.py                  ⬜
+│   └── openai/                                  │   └── openai_api/                    # maps to entrypoints/openai/
+│       ├── api_server.py                        │       ├── test_api_server.py         ⬜  (e2e indirect coverage)
+│       ├── serving_chat.py                       │       ├── test_serving_chat_sampling_params.py  ✅
+│       ├── serving_speech.py                     │       ├── test_serving_speech.py     ✅
+│       └── image_api_utils.py                   │       └── test_image_server.py      ✅
 │
-├── inputs/                    →    ├── inputs/
-│   ├── data.py                     │   ├── test_data.py
-│   ├── parse.py                    │   ├── test_parse.py
-│   └── preprocess.py               │   └── test_preprocess.py
+├── inputs/                             →     ├── inputs/
+│   ├── data.py                                 │   ├── test_data.py                   ⬜
+│   ├── parse.py                                │   ├── test_parse.py                 ⬜
+│   └── preprocess.py                            │   └── test_preprocess.py            ✅ currently in entrypoints/test_omni_input_preprocessor.py
 │
-├── model_executor/            →    ├── model_executor/
-│   ├── layers/                     │   ├── layers/
-│   │   └── mrope.py                │   │   └── test_mrope.py
-│   ├── model_loader/               │   ├── model_loader/
-│   │   └── weight_utils.py         │   │   └── test_weight_utils.py
-│   ├── models/                     │   ├── models/
-│   │   ├── qwen2_5_omni/           │   │   ├── qwen2_5_omni/
-│   │   │   ├── qwen2_5_omni_thinker.py │ │   │   ├── test_qwen2_5_omni_thinker.py  # UT
-│   │   │   ├── qwen2_5_omni_talker.py │ │   │   ├── test_qwen2_5_omni_talker.py  # UT
-│   │   │   └── qwen2_5_omni_token2wav.py │ │   │   └── test_qwen2_5_omni_token2wav.py  # UT
-│   │   └── qwen3_omni/             │   │   └── qwen3_omni/
-│   │       └── ...                 │   │       └── test_*.py
-│   ├── stage_configs/              │   └── stage_configs/             # Configuration tests (if needed)
-│   │   └── ...                     │       └── test_*.py
-│   └── stage_input_processors/     │   └── stage_input_processors/
-│       └── ...                     │       └── test_*.py
+├── model_executor/                     →     ├── model_executor/
+│   ├── layers/                                  │   ├── layers/
+│   │   └── mrope.py                             │   │   └── test_mrope.py              ⬜
+│   ├── model_loader/                            │   ├── model_loader/
+│   │   └── weight_utils.py                      │   │   └── test_weight_utils.py      ⬜
+│   ├── models/                                  │   ├── models/
+│   │   ├── qwen2_5_omni/                         │   │   ├── qwen2_5_omni/
+│   │   │   ├── qwen2_5_omni_thinker.py           │   │   │   ├── test_audio_length.py  ✅
+│   │   │   ├── qwen2_5_omni_talker.py            │   │   │   ├── test_qwen2_5_omni_thinker.py  ⬜
+│   │   │   └── qwen2_5_omni_token2wav.py         │   │   │   ├── test_qwen2_5_omni_talker.py  ⬜
+│   │   └── qwen3_omni/                          │   │   │   └── test_qwen2_5_omni_token2wav.py  ⬜
+│   │       └── ...                               │   │   └── qwen3_omni/
+│   ├── stage_configs/                           │   │       └── test_*.py              ⬜
+│   │   └── *.yaml                               │   └── stage_configs/                 (used by e2e, test_*.py can be added)  ⬜
+│   └── stage_input_processors/                  │   └── stage_input_processors/
+│       └── ...                                  │       └── test_*.py                 ⬜
 │
-├── sample/                    →    ├── sample/
-│   └── ...                         │   └── test_*.py
+├── sample/                             →     ├── sample/
+│   └── __init__.py                             │   └── test_*.py                      ⬜
 │
-├── utils/                     →    ├── utils/
-│   └── platform_utils.py           │   └── test_platform_utils.py
+├── utils/                              →     ├── utils/
+│   └── __init__.py                             │   └── test_*.py                       ⬜  (no platform_utils.py currently)
 │
-├── worker/                    →    ├── worker/
-    ├── gpu_ar_worker.py            │   ├── test_gpu_ar_worker.py
-    ├── gpu_generation_worker.py    │   ├── test_gpu_generation_worker.py
-    ├── gpu_model_runner.py         │   ├── test_gpu_model_runner.py
-    └── npu/                        │   └── npu/                       # Maps to worker/npu/
-        └── ...                     │       └── test_*.py
+├── worker/                             →     ├── worker/
+│   ├── gpu_ar_model_runner.py                  │   ├── test_gpu_ar_model_runner.py    ⬜
+│   ├── gpu_ar_worker.py                        │   ├── test_gpu_ar_worker.py           ⬜
+│   ├── gpu_generation_model_runner.py          │   ├── test_gpu_generation_model_runner.py  ✅
+│   ├── gpu_generation_worker.py                │   ├── test_gpu_generation_worker.py  ⬜
+│   ├── gpu_model_runner.py                     │   ├── test_omni_gpu_model_runner.py   ✅
+│   └── mixins.py                               │   └── (npu under platforms/npu/worker/)  # not worker/npu/
 │
-└── e2e/                       →    ├── e2e/                # End-to-end scenarios (no 1:1 source mirror)
-                                    ├── online_serving/       # Full-stack online serving flows
-                                    │   └── (empty for now)
-                                    └── offline_inference/    # Full offline inference flows
-                                        ├── test_qwen2_5_omni.py     # Moved from multi_stages/
-                                        ├── test_qwen3_omni.py       # Moved from multi_stages_h100/
-                                        ├── test_t2i_model.py  # Moved from single_stage/
-                                        └── stage_configs/           # Shared stage configs
-                                            ├── qwen2_5_omni_ci.yaml
-                                            └── qwen3_omni_ci.yaml
+├── platforms/                          →     (no tests/platforms/, e2e and stage_configs provide indirect coverage)
+│   ├── cuda/
+│   ├── npu/worker/                             # NPU worker here, not vllm_omni/worker/npu/
+│   ├── rocm/
+│   └── xpu/worker/
+│
+├── outputs.py                          →     test_outputs.py                         ✅ (at tests root)
+├── (logger, patch, request, version)    →     (no corresponding unit test)
+│
+└── e2e (tests side only)               →     ├── e2e/
+                                               ├── online_serving/                     ✅ non-empty
+                                               │   ├── test_async_omni.py
+                                               │   ├── test_qwen3_omni.py
+                                               │   ├── test_qwen3_omni_expansion.py
+                                               │   ├── test_image_gen_edit.py
+                                               │   ├── test_images_generations_lora.py
+                                               │   └── stage_configs/
+                                               └── offline_inference/                  ✅
+                                                   ├── test_qwen2_5_omni.py
+                                                   ├── test_qwen3_omni.py
+                                                   ├── test_bagel_text2img.py
+                                                   ├── test_t2i_model.py
+                                                   ├── test_t2v_model.py
+                                                   ├── test_ovis_image.py
+                                                   ├── test_zimage_tensor_parallel.py
+                                                   ├── test_cache_dit.py
+                                                   ├── test_teacache.py
+                                                   ├── test_stable_audio_model.py
+                                                   ├── test_diffusion_cpu_offload.py
+                                                   ├── test_diffusion_layerwise_offload.py
+                                                   ├── test_diffusion_lora.py
+                                                   ├── test_sequence_parallel.py
+                                                   └── stage_configs/
+                                                       ├── qwen2_5_omni_ci.yaml
+                                                       ├── qwen3_omni_ci.yaml
+                                                       ├── bagel_*.yaml
+                                                       └── npu/, rocm/, etc.
 ```
 
 
@@ -117,6 +173,8 @@ vllm_omni/                          tests/
 
 - **E2E Tests**: Place in `tests/e2e/offline_inference/` or `tests/e2e/online_serving/` with descriptive names. Example: `tests/e2e/offline_inference/test_qwen3_omni.py`, `tests/e2e/offline_inference/test_diffusion_model.py`
 
+- **Expansion Tests**
+
 ### Best Practices
 
 1. **Mirror Source Structure**: Test directories should mirror the source code structure
@@ -144,81 +202,217 @@ vllm_omni/                          tests/
 ### Template
 #### E2E - Online serving
 
+E2E Online tests for Qwen3-Omni model with mix input and audio+text output. Based on `tests/e2e/online_serving/test_qwen3_omni.py`.
+
 ```python
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
-Online E2E smoke test for an omni model (video,text,audio → audio).
+E2E Online tests for Qwen3-Omni model with mix input and audio+text output.
 """
+
+import os
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
+
+import threading
 from pathlib import Path
 
-import pytest
 import openai
+import pytest
 
-from tests.utils import hardware_test
+from tests.conftest import (
+    OmniServer,
+    convert_audio_to_text,
+    cosine_similarity_text,
+    dummy_messages_from_mix_data,
+    generate_synthetic_video,
+    merge_base64_and_convert_to_text,
+)
+from vllm_omni.platforms import current_omni_platform
 
-# Optional: set process start method for workers
-os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+# Edit: model name and stage config path
+models = ["Qwen/Qwen3-Omni-30B-A3B-Instruct"]
+
+#If you use the default configuration file, you can directly use the following address.
+def get_default_config():
+    return str(Path(__file__).parent.parent / "stage_configs" / "qwen3_omni_ci.yaml")
+
+#If you need to modify the configuration file, you can use modify_stage_config.
+def get_chunk_config():
+    path = modify_stage_config(
+        get_default_config(),
+        updates={
+            "async_chunk": True,
+            "stage_args": {
+                0: {
+                    "engine_args.custom_process_next_stage_input_func": "vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_async_chunk"
+                },
+                1: {
+                    "engine_args.custom_process_next_stage_input_func": "vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav_async_chunk"
+                },
+            },
+        },
+        deletes={"stage_args": {2: ["custom_process_input_func"]}},
+    )
+    return path
+
+stage_configs = [get_default_config(), CHUNK_CONFIG_PATH]
 
-models = ["{your model name}"] #Edit here to load your model
-stage_configs = [str(Path(__file__).parent / "stage_configs" / {your model yaml})] #Edit here to load your model yaml
 test_params = [(model, stage_config) for model in models for stage_config in stage_configs]
 
-#OmniServer，Used to start the vllm-omni server
-class OmniServer:
-    xxx
+
+#Please use this method to launch the online instance.
+_omni_server_lock = threading.Lock()
+
+@pytest.fixture(scope="module")
+def omni_server(request):
+    """Start vLLM-Omni server as a subprocess. Use module scope for multi-stage init (10-20+ min)."""
+    with _omni_server_lock:
+        model, stage_config_path = request.param
+        with OmniServer(
+            model,
+            ["--stage-configs-path", stage_config_path, "--stage-init-timeout", "120"],
+        ) as server:
+            yield server
 
 
 @pytest.fixture
-def omni_server(request):
-    model, stage_config_path = request.param
-    with OmniServer(model, ["--stage-configs-path", stage_config_path]) as server:
-        yield server
-
-
-#handle request message
-@pytest.fixture(scope="session")
-def base64_encoded_video() -> str:
-    xxx
-
-@pytest.fixture(scope="session")
-def dummy_messages_from_video_data(video_data_url: str, content_text: str) -> str:
-    xxx
-
-@pytest.mark.core_model
-@pytest.mark.omni
-@hardware_test(
-    res={"cuda": "L4", "rocm": "MI325", "npu": "A2"},
-    num_cards={"cuda": 2, "rocm": 2, "npu": 4},
-)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
-def test_video_to_audio(
-    client: openai.OpenAI,
-    omni_server,
-    base64_encoded_video: str,
-) -> None:
-    #set message
-    video_data_url = f"data:video/mp4;base64, {base64_encoded_video}"
-    messages = dummy_messages_from_video_data(video_data_url)
-
-    #send request
-    chat_completion = client.chat.completions.create(
-        model=omni_server.model,
-        messages=messages,
+def client(omni_server):
+    """OpenAI client for the running vLLM-Omni server."""
+    return openai.OpenAI(
+        base_url=f"http://{omni_server.host}:{omni_server.port}/v1",
+        api_key="EMPTY",
     )
 
-    #verify text output
-    text_choice = chat_completion.choices[0]
-    assert text_choice.finish_reason == "length"
+#Please use function definitions above the test function to define the prompts and other parameters you need.
+def get_system_prompt():
+    return {
+        "role": "system",
+        "content": [
+            {
+                "type": "text",
+                "text": (
+                    "You are Qwen, a virtual human developed by the Qwen Team, "
+                    "Alibaba Group, capable of perceiving auditory and visual inputs, "
+                    "as well as generating text and speech."
+                ),
+            }
+        ],
+    }
 
-    #verify audio output
-    audio_choice = chat_completion.choices[1]
-    audio_message = audio_choice.message
-    if hasattr(audio_message, "audio") and audio_message.audio:
-        assert audio_message.audio.data is not None
-        assert len(audio_message.audio.data) > 0
+...
+
+#Please define test case tags according to the instructions in the marker documentation.
+@pytest.mark.core_model
+@pytest.mark.omni
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_mix_to_text_audio_001(client: openai.OpenAI, omni_server, request) -> None:
+    # PLEASE FOLLOW THESE TEMPLATE INSTRUCTIONS:
+    # ============================================================================
+    # TEMPLATE USAGE GUIDE:
+    # 1. Copy this entire function as a starting point for multi-modal tests
+    # 2. Update the test name to reflect your specific test scenario
+    # 3. Modify input/output modalities as needed (see OPTIONS section below)
+    # 4. Adjust assertions based on your expected outcomes
+    # 5. Add custom validation logic for your specific use case
+    # ============================================================================
+
+    #Please list the relevant test points.
+    """
+    Test multi-modal input processing and text/audio output generation via OpenAI API.
+    Deploy Setting: default yaml
+    Input Modal: text + audio + video + image
+    Output Modal: text + audio
+    Input Setting: stream=True
+    Datasets: single request
+    """
+    # SECTION 1: TEST SETUP AND INITIALIZATION
+    # =========================================
+    # INSTRUCTIONS: Initialize test variables and prepare test environment
+    # MODIFY: Add any additional test setup required for your scenario
+    e2e_list = list()
+    # SECTION 2: TEST DATA GENERATION
+    # ================================
+    # INSTRUCTIONS: Generate or load test data for each input modality
+    # MODIFY: Replace synthetic generators with your actual data sources
+    # VIDEO DATA - Generate synthetic video for testing
+    # FORMAT: data:video/mp4;base64,{base64_encoded_video}
+    # PARAMETERS: width, height, duration_frames
+    video_data_url = f"data:video/mp4;base64,{generate_synthetic_video(224, 224, 300)['base64']}"
+    # IMAGE DATA - Generate synthetic image for testing
+    # FORMAT: data:image/jpeg;base64,{base64_encoded_image}
+    # PARAMETERS: width, height
+    image_data_url = f"data:image/jpeg;base64,{generate_synthetic_image(224, 224)['base64']}"
+    # AUDIO DATA - Generate synthetic audio for testing
+    # FORMAT: data:audio/wav;base64,{base64_encoded_audio}
+    # PARAMETERS: duration_seconds, channels
+    audio_data_url = f"data:audio/wav;base64,{generate_synthetic_audio(5, 1)['base64']}"
+
+    # SECTION 3: MESSAGE CONSTRUCTION
+    # ================================
+    # INSTRUCTIONS: Assemble the complete message payload for API request
+    # MODIFY: Add/remove modalities or change prompt structure as needed
+
+    # USAGE: Construct a message containing all input modalities
+    # IMPORTANT: Ensure the message structure matches OpenAI API expectations
+    # CUSTOMIZATION POINTS:
+    #   - system_prompt: Controls the assistant's behavior
+    #   - content_text: The user's text prompt/question
+    #   - *_data_url: URLs for media content (video/image/audio)
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        video_data_url=video_data_url,
+        image_data_url=image_data_url,
+        audio_data_url=audio_data_url,
+        content_text=get_prompt("mix"),
+    )
+
+    # SECTION 4: API REQUEST EXECUTION
+    # =================================
+    # INSTRUCTIONS: Make the API call and measure performance
+    # MODIFY: Add timeout, retry logic, or additional parameters
+    start_time = time.perf_counter()
+    chat_completion = client.chat.completions.create(model=omni_server.model, messages=messages, stream=True)
+
+    #Call using your preferred method and obtain the final audio and text outputs.
+    ...
+
+    # SECTION 5: OUTPUT VALIDATION
+    # =============================
+    # INSTRUCTIONS: Verify that outputs meet expected criteria
+    # MODIFY: Adjust validation logic for your specific requirements
+
+    # ASSERTION 1: E2E Validation
+    # PURPOSE: Verify that the E2E latency is less than the baseline.
+    current_e2e = time.perf_counter() - start_time
+    print(f"the request e2e is: {current_e2e}")
+    e2e_list.append(current_e2e)
+
+    print(f"the avg e2e is: {sum(e2e_list) / len(e2e_list)}")
+
+
+
+    # ASSERTION 2: Text Output Validation
+    # PURPOSE: Verify that text output was generated with keyword content
+    assert text_content is not None and len(text_content) >= 2, "No text output is generated"
+    assert any(
+        keyword in text_content.lower() for keyword in ["square", "quadrate", "sphere", "globe", "circle", "round"]
+    ), "The output does not contain any of the keywords."
+
+
+    # ASSERTION 3: Cross-Modal Consistency
+    # PURPOSE: Verify text and audio outputs convey the same information
+    # CUSTOMIZATION: Adjust similarity threshold (0.9) based on accuracy requirements
+    assert audio_data is not None, "No audio output is generated"
+    audio_content = merge_base64_and_convert_to_text(audio_data)
+    print(f"text content is: {text_content}")
+    print(f"audio content is: {audio_content}")
+    similarity = cosine_similarity_text(audio_content.lower(), text_content.lower())
+    print(f"similarity is: {similarity}")
+    assert similarity > 0.9, "The audio content is not same as the text"
 ```
 
+
 #### E2E - Offline inference
 ```python
 # SPDX-License-Identifier: Apache-2.0
@@ -271,10 +465,3 @@ def test_video_to_audio(omni_runner: type[OmniRunner], model: str) -> None:
         has_audio = any(o.final_output_type == "audio" for o in outputs)
         assert has_audio
 ```
-
-## Checklist before submitting your test files
-
-1. The file is saved in an appropriate place and the file name is clear.
-2. The coding style follows the requirements outlined above.
-3. **All test functions have appropriate pytest markers**
-4. For tests that need run in CI, please ensure the test is configured under the `./buildkite/` folder.

docs/contributing/metrics.md

@@ -0,0 +1,173 @@
+
+# Metrics
+
+You can use these metrics in production to monitor the health and performance of the vLLM-omni system. Typical scenarios include:
+
+- **Performance Monitoring**: Track throughput (e.g., `e2e_avg_tokens_per_s`), latency (e.g., `e2e_total_ms`), and resource utilization to verify that the system meets expected standards.
+
+- **Debugging and Troubleshooting**: Use detailed per-request metrics to diagnose issues, such as high transfer times or unexpected token counts.
+
+## How to Enable and View Metrics
+
+### Start the Service with Metrics Logging
+
+```bash
+vllm serve /workspace/models/Qwen3-Omni-30B-A3B-Instruct --omni --port 8014 --log-stats
+```
+
+### Send a Request
+
+```bash
+python openai_chat_completion_client_for_multimodal_generation.py --query-type use_image
+```
+
+### What You Will See
+
+With `--log-stats` enabled, the server will output detailed metrics logs after each request. Example output:
+
+
+#### Overall Summary
+
+| Field                       | Value        |
+|-----------------------------|--------------|
+| e2e_requests                | 1            |
+| e2e_wall_time_ms            | 41,299.190   |
+| e2e_total_tokens            | 5,202        |
+| e2e_avg_time_per_request_ms | 41,299.190   |
+| e2e_avg_tokens_per_s        | 125.959      |
+| e2e_stage_0_wall_time_ms    | 10,192.289   |
+| e2e_stage_1_wall_time_ms    | 30,541.409   |
+| e2e_stage_2_wall_time_ms    |    207.496   |
+
+#### RequestE2EStats
+
+| Field                   | Value      |
+|-------------------------|------------|
+| e2e_total_ms            | 41,299.133 |
+| e2e_total_tokens        | 5,202      |
+| transfers_total_time_ms | 245.895    |
+| transfers_total_kbytes  | 138,089.939|
+
+#### StageRequestStats
+
+| Field                  | 0      | 1      | 2      |
+|------------------------|--------|--------|--------|
+| audio_generated_frames | 0      | 0      | 525,525|
+| batch_id               | 38     | 274    | 0      |
+| batch_size             | 1      | 1      | 1      |
+| num_tokens_in          | 4,860  | 4,826  | 4,384  |
+| num_tokens_out         | 67     | 275    | 0      |
+| postprocess_time_ms    | 256.158| 0.491  | 0.000  |
+| stage_gen_time_ms      | 9,910.007|30,379.198|160.745|
+
+#### TransferEdgeStats
+
+| Field               | 0->1        | 1->2       |
+|---------------------|-------------|------------|
+| size_kbytes         | 109,277.349 | 28,812.591 |
+| tx_time_ms          | 78.701      | 18.790     |
+| rx_decode_time_ms   | 111.865     | 31.706     |
+| in_flight_time_ms   | 2.015       | 2.819      |
+
+
+These logs include:
+
+- **Overall summary**: total requests, wall time, average tokens/sec, etc.
+
+- **E2E table**: per-request latency and token counts.
+
+- **Stage table**: per-stage batch and timing details.
+
+- **Transfer table**: data transfer and timing for each edge.
+
+You can use these logs to monitor system health, debug performance, and analyze request-level metrics as described above.
+
+
+## Metrics Scope: Offline vs Online Inference
+
+For **offline inference** (batch mode), the summary includes both system-level metrics (aggregated across all requests) and per-request metrics. In this case, `e2e_requests` can be greater than 1, reflecting multiple completed requests in a batch.
+
+For **online inference** (serving mode), the summary is always per-request. `e2e_requests` is always 1, and only request-level metrics are reported for each completion.
+
+---
+
+## Parameter Details
+
+### Summary Metrics
+
+| Field                     | Meaning                                                                                       |
+|---------------------------|----------------------------------------------------------------------------------------------|
+| `e2e_requests`            | Number of completed requests.                                                                |
+| `e2e_wall_time_ms`        | Wall-clock time span from run start to last completion, in ms.                               |
+| `e2e_total_tokens`        | Total tokens counted across all completed requests (stage0 input + all stage outputs).       |
+| `e2e_avg_time_per_request_ms` | Average wall time per request: `e2e_wall_time_ms / e2e_requests`.                        |
+| `e2e_avg_tokens_per_s`    | Average token throughput over wall time: `e2e_total_tokens * 1000 / e2e_wall_time_ms`.      |
+| `e2e_stage_{i}_wall_time_ms` | Wall-clock time span for stage i, in ms. Each stage's wall time is reported as a separate field, e.g., `e2e_stage_0_wall_time_ms`, `e2e_stage_1_wall_time_ms`, etc. |
+
+---
+
+### E2E Table (per request)
+
+| Field                     | Meaning                                                               |
+|---------------------------|-----------------------------------------------------------------------|
+| `e2e_total_ms`            | End-to-end latency in ms.                                             |
+| `e2e_total_tokens`        | Total tokens for the request (stage0 input + all stage outputs).      |
+| `transfers_total_time_ms` | Sum of transfer edge `total_time_ms` for this request.                |
+| `transfers_total_kbytes`  | Sum of transfer kbytes for this request.                              |
+
+
+---
+
+### Stage Table (per stage event / request)
+
+| Field                     | Meaning                                                                                         |
+|---------------------------|-------------------------------------------------------------------------------------------------|
+| `batch_id`                | Batch index.                                                                                    |
+| `batch_size`              | Batch size.                                                                                     |
+| `num_tokens_in`           | Input tokens to the stage.                                                                      |
+| `num_tokens_out`          | Output tokens from the stage.                                                                   |
+| `stage_gen_time_ms`       | Stage compute time in ms, excluding postprocessing time (reported separately as `postprocess_time_ms`). |
+| `image_num`               | Number of images generated (for diffusion/image stages).                                        |
+| `resolution`              | Image resolution (for diffusion/image stages).                                                                  |
+| `postprocess_time_ms` | Diffusion/image: post-processing time in ms.                                                    |
+
+---
+
+### Transfer Table (per edge / request)
+
+| Field                | Meaning                                                                   |
+|----------------------|---------------------------------------------------------------------------|
+| `size_kbytes`        | Total kbytes transferred.                                                 |
+| `tx_time_ms`         | Sender transfer time in ms.                                               |
+| `rx_decode_time_ms`  | Receiver decode time in ms.                                               |
+| `in_flight_time_ms`  | In-flight time in ms.                                                     |
+
+
+### Expectation of the Numbers (Verification)
+
+**Formulas:**
+
+- `e2e_total_tokens = Stage0's num_tokens_in + sum(all stages' num_tokens_out)`
+
+- `transfers_total_time_ms = sum(tx_time_ms + rx_decode_time_ms + in_flight_time_ms)` for every edge
+
+**Using the example above:**
+
+**e2e_total_tokens**
+
+- Stage0's `num_tokens_in`: **4,860**
+- Stage0's `num_tokens_out`: **67**
+- Stage1's `num_tokens_out`: **275**
+- Stage2's `num_tokens_out`: **0**
+
+so `e2e_total_tokens = 4,860 + 67 + 275 + 0 = 5,202`, which matches the table value `e2e_total_tokens`.
+
+**transfers_total_time_ms**
+
+For each edge:
+
+- 0->1: tx_time_ms (**78.701**) + rx_decode_time_ms (**111.865**) + in_flight_time_ms (**2.015**) = **192.581**
+
+- 1->2: tx_time_ms (**18.790**) + rx_decode_time_ms (**31.706**) + in_flight_time_ms (**2.819**) = **53.315**
+
+192.581 + 53.315 = **245.896** = transfers_total_time_ms, which matches the calculation (difference is due to rounding)

docs/contributing/profiling.md

@@ -106,7 +106,7 @@ python image_to_video.py \
     #    Reducing frames to the absolute minimum (2) keeps the
     #    tensor size small, ensuring the trace file doesn't become
     #    multi-gigabytes in size.
-    --num_frames 2 \
+    --num-frames 2 \
     \
     # Minimize Iteration Loop (Steps):
     #    This is the most critical setting for profiling.
@@ -114,12 +114,12 @@ python image_to_video.py \
     #    Profiling 2 steps gives you the exact same performance
     #    data as 50 steps, but saves minutes of runtime and
     #    prevents the trace viewer from freezing.
-    --num_inference_steps 2 \
+    --num-inference-steps 2 \
     \
-    --guidance_scale 5.0 \
-    --guidance_scale_high 6.0 \
-    --boundary_ratio 0.875 \
-    --flow_shift 12.0 \
+    --guidance-scale 5.0 \
+    --guidance-scale-high 6.0 \
+    --boundary-ratio 0.875 \
+    --flow-shift 12.0 \
     --fps 16 \
     --output i2v_output.mp4
 
@@ -131,9 +131,6 @@ python image_to_video.py \
 
 2. **Wan-AI/Wan2.2-I2V-A14B-Diffusers**:   [https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_video](https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_video)
 
-> **Note:**
-As of now, asynchronous (online) profiling is not fully supported in vLLM-Omni. While start_profile() and stop_profile() methods exist, they are only reliable in offline inference scripts (e.g., the provided end2end.py examples). Do not use them in server-mode or streaming scenarios—traces may be incomplete or fail to flush.
-
 ### 4. Analyzing Omni Traces
 
 Output files are saved to your configured ```VLLM_TORCH_PROFILER_DIR```.

docs/design/feature/async_chunk_design.md

@@ -0,0 +1,209 @@
+# Async Chunk Design
+
+## Table of Contents
+
+1. [Overview](#overview)
+2. [Performance](#performance)
+3. [Architecture](#architecture)
+4. [Configuration](#configuration)
+5. [Related Files](#related-files)
+
+## Overview
+
+The `async_chunk` feature enables asynchronous, chunked processing of data across multiple stages in a multi-stage pipeline (e.g., Qwen3-Omni with Thinker → Talker → Code2Wav stages). Instead of waiting for a complete stage output before forwarding to the next stage, this feature allows stages to process and forward data in chunks as it becomes available, significantly reducing latency and improving throughput.
+
+**Chunk Size Definition**
+
+- **Prefill Phase**: `chunk_size = num_scheduled_tokens` for chunked prefill processing
+- **Decode Phase**: `chunk_size = num_scheduled_tokens = 1 ` for per-token streaming
+
+For qwen3-omni:
+- **Thinker → Talker**: Per decode step (typically chunk_size=1)
+- **Talker → Code2Wav**: Accumulated to code2wav chunk_size(default=25, currently only support default, will support chunk_size soon) before sending
+- **Code2Wav**: Streaming decode with code2wav chunk_size
+
+With `async_chunk`:
+- Stages can start processing as soon as chunks are available
+- Overlapping execution across stages
+- Reduced latency and improved throughput
+- Better resource utilization
+- Async scheduling: Chunk IO (get/put) overlaps with compute via background threads so the scheduler is not blocked waiting for chunks
+
+## Performance
+1. **Reduced Latency**: Next stage can start processing immediately
+2. **Streaming Support**: Enables streaming for audio generation
+3. **IO-Compute Overlap**: Chunk retrieval happens asynchronously while other requests compute
+4. **Non-blocking Scheduler**: Requests waiting for chunks don't block the entire scheduler
+5. **Code2Wav Batch Inference**: Supports batched processing in code2wav stage
+
+| Input     | Output           | Async_chunk enabled | Code2Wav batch size | Max_Concurrency | Prompts | Mean E2E   | Mean TTFT  | Mean TPOT | Mean TTFP   | Mean RTF | Mean ITL |
+|-----------|------------------|---------------------|---------------------|----------------|---------|------------|------------|-----------|-------------|----------|----------|
+| text 100  | text 100+audio   | False               | 1                   | 1              | 50      | 6581.80    | 43.22      | 8.31      | 6459.34     | 0.24     | 8.22     |
+| text 100  | text 100+audio   | False               | 1                   | 4              | 50      | 7398.63    | 67.57      | 9.14      | 7285.35     | 0.27     | 9.05     |
+| text 100  | text 100+audio   | False               | 1                   | 10             | 50      | 13522.99   | 131.82     | 12.72     | 13410.44    | 0.49     | 12.60    |
+| text 100  | text 100+audio   | False               | 64                  | 1              | 50      | 6505.13    | 43.14      | 8.52      | 6395.40     | 0.24     | 8.44     |
+| text 100  | text 100+audio   | False               | 64                  | 4              | 50      | 7668.15    | 51.15      | 9.36      | 7562.37     | 0.28     | 9.27     |
+| text 100  | text 100+audio   | False               | 64                  | 10             | 50      | 9516.18    | 138.06     | 14.75     | 9409.26     | 0.34     | 14.60    |
+| text 100  | text 100+audio   | True                | 1                   | 1              | 50      | 6179.79    | 44.58      | 8.69      | 522.99      | 0.22     | 8.60     |
+| text 100  | text 100+audio   | True                | 1                   | 4              | 50      | 7692.69    | 103.96     | 10.22     | 785.85      | 0.29     | 10.12    |
+| text 100  | text 100+audio   | True                | 1                   | 10             | 50      | 11152.71   | 685.60     | 17.64     | 1628.88     | 0.41     | 17.62    |
+
+
+Performance data collected on H800 GPUs through comprehensive benchmarking with cudagraph enabled. text input uses random dataset.
+
+Enabling **async_chunk** (False→True) sharply reduces time-to-first-audio (TTFP)—e.g. ~92% at concurrency 1 (6.5s→0.52s)—and improves E2E latency (e.g. ~6% at conc 1, ~17% at conc 10). RTF (Real Time Factor) also improves with async_chunk on (e.g. ~8% at conc 1: 0.24→0.22, ~16% at conc 10: 0.49→0.41). Enabling **Code2Wav batch size 64** (vs 1) improves E2E and TTFP at higher concurrency when async_chunk is off (e.g. ~30% at conc 10: 13.5s→9.5s E2E, 13.4s→9.4s TTFP).
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/performance/qwen3-omni_ttfp_performance.png">
+    <img alt="TTFP Performance Data Comparison" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/performance/qwen3-omni_ttfp_performance.png" width=100%>
+  </picture>
+</p>
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/performance/qwen3-omni_rtf_performance.png">
+    <img alt="RTF Performance Data Comparison" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/performance/qwen3-omni_rtf_performance.png" width=100%>
+  </picture>
+</p>
+
+## Architecture
+### Data Flow
+
+#### Sequential Flow
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/qwen3-omni-non-async-chunk.png">
+    <img alt="Data Flow between stages" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/qwen3-omni-non-async-chunk.png" width=100%>
+  </picture>
+</p>
+
+#### Async Chunk Flow
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/qwen3-omni-async-chunk.png">
+    <img alt="Data Flow between stages" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/qwen3-omni-async-chunk.png" width=100%>
+  </picture>
+</p>
+
+### Async Chunk architecture
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/async-chunk-architecture.png">
+    <img alt="Async Chunk Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/async-chunk-architecture.png" width=100%>
+  </picture>
+</p>
+
+
+### Key Components
+
+1. **OmniConnector**: Inter-stage data transport only
+   - Shared memory or other IPC mechanisms
+   - **Transport-only API**: `put(from_stage, to_stage, put_key, data)` and `get(from_stage, to_stage, get_key)` (optionally with timeout)
+   - **No request-specific state**: Connector does not track put_requests, get_requests, request_payload, finished_requests, or other request-bound metadata; it only performs put/get operations
+   - Chunk keys and request/chunk lifecycle are managed by **OmniChunkTransferAdapter**
+
+2. **Transfer Adapter Layer**: Extensible abstraction for managing data transfer via connectors
+   - **OmniTransferAdapterBase**: Base class with background **recv_loop** and **save_loop** threads;
+   - **OmniChunkTransferAdapter**: Chunk-specific implementation that owns the full chunk lifecycle when async_chunk is enabled
+     - **Chunk ID and key construction**: Builds keys like `{req_id}_{stage_id}_{chunk_id}` for put/get
+     - **Async get**: `load_async(request)` enqueues the request; background **recv_loop** polls the connector (non-blocking); when data is available, updates the request and marks it in `_finished_load_reqs`; scheduler calls `get_finished_requests()` to learn which requests have chunks ready
+     - **Async put**: `save_async(pooling_output, request)` invokes `custom_process_next_stage_input_func` in the main thread to build the payload, then enqueues a save task; background **save_loop** performs `connector.put()`; payload processing and chunk accumulation (e.g. code2wav chunk_size) remain in the main thread
+
+3. **Stage Input Processors**: Custom functions that process stage outputs into chunks for different models
+   - Receive **transfer_manager** (OmniChunkTransferAdapter)
+   - Qwen3-omni reference: `thinker2talker_async_chunk`, `talker2code2wav_async_chunk`
+
+4. **Schedulers**: Modified to handle chunk-based scheduling with async IO-compute overlap
+   - `OmniARScheduler`: For autoregressive stages
+   - `OmniGenerationScheduler`: For generation stages
+   - Both schedulers use **OmniChunkTransferAdapter** and **before/after** hooks around `super().schedule()`:
+     - **Before** `super().schedule()`: `process_pending_chunks(waiting, running)` moves requests waiting for chunks to `WAITING_FOR_CHUNK`, enqueues load tasks for background polling
+     - **After** `super().schedule()`: `restore_queues(waiting, running)` restores requests with ready chunks back to waiting/running, `postprocess_scheduler_output(scheduler_output)` attaches cached additional_information, clears chunk-ready flags
+   - **put_chunk** `save_async(pooler_output, request)`; **get_chunk** / **get_chunk_for_generation** `load_async(request)`
+
+5. **Model Runners**: Handle chunk processing
+   - `OmniGPUModelRunner`: Processes chunks in AR stages
+   - `GPUGenerationModelRunner`: Processes chunks in generation stages
+     - Uses `ubatch_slices` from `get_forward_context()` to track per-request sequence lengths in batched inference
+     - Reuses `ubatch_slices_padded` for code2wav batching to properly split batch outputs
+     - Handles list-type multimodal outputs: iterates through requests and assigns corresponding tensor to each
+     - Improved request state management: removes unscheduled and finished requests from input batch
+
+6. **Model Implementation**: Model-specific chunk handling
+   - `Qwen3OmniMoeForConditionalGeneration`: Main model with async_chunk support
+     - **Code2Wav stage batching**: Uses `ubatch_slices` to construct batched codec codes tensor `[batch_size, 16, max_seq_len]`
+     - **Batch output handling**: `generate_audio()` returns `list[torch.Tensor]`, one audio tensor per request
+     - **Multimodal outputs**: Returns list of audio tensors for batch processing instead of single concatenated tensor
+   - `Qwen3OmniCode2WavDecoder`: Audio generation model
+     - `chunked_decode()` and `chunked_decode_streaming()`: Return `list[torch.Tensor]` (one per request)
+     - Uses `ubatch_slices` to split batched waveform output into per-request audio chunks
+     - Each request gets correctly sized audio based on its code sequence length: `waveform[:, :, :code_seq_len * total_upsample]`
+
+7. **Request status**: `RequestStatus.WAITING_FOR_CHUNK` is added via patch (e.g. in `vllm_omni/patch.py`) so requests waiting for a chunk are not scheduled by the base vLLM scheduler until the chunk is ready.
+
+## Configuration
+
+Enable async_chunk in stage configuration YAML:
+
+```yaml
+async_chunk: true
+stage_args:
+  - stage_id: 0
+    engine_args:
+      custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_async_chunk
+  - stage_id: 1
+    engine_args:
+      custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav_async_chunk
+```
+
+### Stage Configuration
+
+- `async_chunk: bool`: Enable/disable async chunk mode
+- `custom_process_next_stage_input_func: str`: Path to custom chunk processing function; receives `(transfer_manager, pooling_output, request)`. For qwen3-omni: `thinker2talker_async_chunk`, `talker2code2wav_async_chunk`
+- `stage_connector_config: dict`: Connector configuration
+- `worker_type: str`: Model type, e.g. `"ar"` or `"generation"` (used by OmniChunkTransferAdapter for mode-specific payload handling)
+- `max_batch_size: int`: Maximum batch size for the stage
+
+
+### Connector Configuration
+
+```yaml
+connectors:
+  - from_stage: 0
+    to_stage: 1
+    spec:
+      name: SharedMemoryConnector
+      extra:
+        stage_id: 0
+```
+
+### Code2Wav Batch Configuration
+
+For optimal performance with async_chunk, the code2wav stage should be configured with batching:
+
+```yaml
+stage_args:
+  - stage_id: 2  # code2wav stage
+    runtime:
+      devices: "1"
+      max_batch_size: 64  # Enables batched audio generation
+    engine_args:
+      model_stage: code2wav
+```
+
+## Related Files
+
+- `vllm_omni/model_executor/stage_input_processors/qwen3_omni.py`: Chunk processing functions (receive `transfer_manager` as first param)
+- `vllm_omni/distributed/omni_connectors/transfer_adapter/base.py`: OmniTransferAdapterBase (recv_loop, save_loop, load_async, save_async)
+- `vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py`: OmniChunkTransferAdapter (process_pending_chunks, restore_queues, postprocess_scheduler_output)
+- `vllm_omni/distributed/omni_connectors/connectors/shm_connector.py`: SharedMemoryConnector (transport-only put/get)
+- `vllm_omni/core/sched/omni_ar_scheduler.py`: AR scheduler with chunk_transfer_adapter
+- `vllm_omni/core/sched/omni_generation_scheduler.py`: Generation scheduler with same async chunk pattern
+- `vllm_omni/worker/gpu_model_runner.py`: Model runner with chunk handling
+- `vllm_omni/worker/gpu_generation_model_runner.py`: Generation model runner with batch output handling and ubatch_slices support
+- `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py`: Model implementation with code2wav batching
+- `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni_code2wav.py`: Code2wav decoder with batch support
+- `vllm_omni/engine/arg_utils.py`: Configuration definitions (async_chunk, worker_type)
+- `vllm_omni/config/model.py`: Model config with async_chunk field

docs/design/feature/cache_dit.md

@@ -0,0 +1,286 @@
+# Cache-DiT
+
+This section describes how to add cache-dit acceleration to a new diffusion pipeline. We use the Qwen-Image pipeline and LongCat-Image pipeline as reference implementations.
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Standard Models: Automatic Support](#standard-models-automatic-support)
+- [Custom Architectures: Writing Custom Implementation](#custom-architectures-writing-custom-implementation)
+- [Testing](#testing)
+- [Troubleshooting](#troubleshooting)
+- [Reference Implementations](#reference-implementations)
+- [Summary](#summary)
+
+---
+
+## Overview
+
+### What is Cache-DiT?
+
+Cache-DiT is an acceleration library for Diffusion Transformers (DiT) that caches intermediate computation results across denoising steps. The core insight is that adjacent denoising steps often produce similar intermediate features, so we can skip redundant computations by reusing cached results.
+
+The library supports three main caching strategies:
+
+- **DBCache:** Dynamic block-level caching that selectively computes or caches transformer blocks based on residual differences
+- **TaylorSeer:** Calibration-based prediction that estimates block outputs using Taylor expansion
+- **SCM (Step Computation Masking):** Dynamic step skipping based on configurable policies
+
+### Architecture
+
+vLLM-omni integrates cache-dit through the `CacheDiTBackend` class, which provides a unified interface for managing cache-dit acceleration on diffusion models.
+
+| Method/Class | Purpose | Behavior |
+|--------------|---------|----------|
+| [`CacheDiTBackend`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/#vllm_omni.diffusion.cache.CacheBackend) | Unified backend interface | Automatically handles enabler selection and cache refresh |
+| [`enable_cache_for_dit()`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/cache_dit_backend/#vllm_omni.diffusion.cache.cache_dit_backend.enable_cache_for_dit) | Apply caching to transformer | Configures DBCache on transformer blocks |
+
+**Key APIs from Cache-DiT:**
+
+[Cache-DiT API Reference](https://cache-dit.readthedocs.io/en/latest/user_guide/CACHE_API/)
+
+| API | Description |
+|-----|-------------|
+| `BlockAdapter` | Core abstraction for applying cache-dit to transformers. Specifies transformer module(s), block list(s), and forward signature pattern(s). |
+| `ForwardPattern` | Defines block forward signature patterns: `Pattern_0`, `Pattern_1`, `Pattern_2` |
+| `ParamsModifier` | Per-transformer or per-block-list cache configuration customization |
+| `DBCacheConfig` | Configuration for DBCache parameters (warmup steps, cached steps, thresholds) |
+| `refresh_context()` | Update cache context | Called when `num_inference_steps` changes |
+
+---
+
+## Standard Models: Automatic Support
+
+Most DiT models follow this pattern:
+- Single transformer with one `ModuleList` of blocks
+- Standard forward signature
+- Compatible with cache-dit's automatic detection
+
+**Examples:** Qwen-Image, Z-Image
+
+For standard single-transformer models, **no code changes are needed**. The `CacheDiTBackend` automatically uses `enable_cache_for_dit()`:
+
+```python
+from vllm_omni import Omni
+
+# Works automatically for standard models
+omni = Omni(
+    model="Qwen/Qwen-Image",  # Standard single-transformer model
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+    }
+)
+```
+
+**What happens automatically:**
+
+```python
+def enable_cache_for_dit(pipeline: Any, cache_config: Any) -> Callable[[int], None]:
+    """Default enabler for standard single-transformer DiT models."""
+
+    # Build cache configuration
+    db_cache_config = DBCacheConfig(
+        num_inference_steps=None,  # Will be set during first inference
+        Fn_compute_blocks=cache_config.Fn_compute_blocks,
+        Bn_compute_blocks=cache_config.Bn_compute_blocks,
+        max_warmup_steps=cache_config.max_warmup_steps,
+        max_cached_steps=cache_config.max_cached_steps,
+        max_continuous_cached_steps=cache_config.max_continuous_cached_steps,
+        residual_diff_threshold=cache_config.residual_diff_threshold,
+    )
+
+    # Enable cache-dit on transformer
+    cache_dit.enable_cache(
+        pipeline.transformer,
+        cache_config=db_cache_config,
+    )
+
+    # Return refresh function for dynamic num_inference_steps updates
+    def refresh_cache_context(pipeline: Any, num_inference_steps: int, verbose: bool = True):
+        cache_dit.refresh_context(pipeline.transformer, num_inference_steps=num_inference_steps, verbose=verbose)
+
+    return refresh_cache_context
+```
+
+---
+
+## Custom Architectures: Writing Custom Implementation
+
+Some models require custom handling:
+
+- **Dual-transformer:** Models with separate high-noise and low-noise transformers (e.g., Wan2.2)
+- **Multi-block-list:** Models with multiple block lists in one transformer (e.g., LongCatImage with `transformer_blocks` + `single_transformer_blocks`)
+- **Special forward patterns:** Models with non-standard block execution patterns
+
+### Example 1: Dual-Transformer Model (Wan2.2)
+
+Wan2.2 uses two transformers: one for high-noise steps and one for low-noise steps.
+
+**Key difference:** Use `BlockAdapter` to wrap multiple transformers with separate configurations.
+
+```python
+# Standard: cache_dit.enable_cache(pipeline.transformer, ...)
+# Custom: Use BlockAdapter to handle multiple transformers
+cache_dit.enable_cache(
+    BlockAdapter(
+        transformer=[pipeline.transformer, pipeline.transformer_2],  # Multiple transformers
+        blocks=[pipeline.transformer.blocks, pipeline.transformer_2.blocks],
+        forward_pattern=[ForwardPattern.Pattern_2, ForwardPattern.Pattern_2],
+        params_modifiers=[
+            ParamsModifier(...),  # Config for high-noise transformer
+            ParamsModifier(...),  # Config for low-noise transformer (different params)
+        ],
+    ),
+    cache_config=db_cache_config,
+)
+```
+
+**Key difference:** `refresh_context` must be called on each transformer separately.
+
+```python
+# Standard: cache_dit.refresh_context(pipeline.transformer, num_inference_steps=N)
+# Custom: Refresh each transformer with its own step count
+def refresh_cache_context(pipeline, num_inference_steps, verbose=True):
+    high_steps, low_steps = _split_inference_steps(num_inference_steps)
+    cache_dit.refresh_context(pipeline.transformer, num_inference_steps=high_steps, ...)
+    cache_dit.refresh_context(pipeline.transformer_2, num_inference_steps=low_steps, ...)
+```
+
+### Example 2: Multi-Block-List Model (LongCatImage)
+
+LongCatImage has a single transformer with two block lists: `transformer_blocks` and `single_transformer_blocks`.
+
+**Key difference:** Use `BlockAdapter` to specify multiple block lists within one transformer.
+
+```python
+# Standard: cache_dit.enable_cache(pipeline.transformer, ...)
+#   - Automatically detects single block list
+# Custom: Use BlockAdapter to specify multiple block lists
+cache_dit.enable_cache(
+    BlockAdapter(
+        transformer=pipeline.transformer,  # Single transformer
+        blocks=[
+            pipeline.transformer.transformer_blocks,        # Block list 1
+            pipeline.transformer.single_transformer_blocks, # Block list 2
+        ],
+        forward_pattern=[ForwardPattern.Pattern_1, ForwardPattern.Pattern_1],
+        params_modifiers=[modifier],
+    ),
+    cache_config=db_cache_config,
+)
+```
+
+> **Note:** For single transformer with multiple block lists, `refresh_context` works the same as standard models.
+
+### Registering Custom Implementations
+
+After writing your custom enabler, register it in `CUSTOM_DIT_ENABLERS` in `vllm_omni/diffusion/cache/cache_dit_backend.py`:
+
+```python
+CUSTOM_DIT_ENABLERS = {
+    "Wan22Pipeline": enable_cache_for_wan22,
+    "LongCatImagePipeline": enable_cache_for_longcat_image,
+    "YourCustomPipeline": enable_cache_for_your_model,  # Add here
+}
+```
+
+---
+
+## Testing
+
+After adding cache-dit support, test with:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+# Test your custom model
+omni = Omni(
+    model="your-model-name",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+        "residual_diff_threshold": 0.24,
+    }
+)
+
+images = omni.generate(
+    "a beautiful landscape",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+**Verify:**
+
+1. Cache is applied (check logs for "Cache-dit enabled successfully on xxx")
+2. Performance improvement (should be around 1.5x-2x faster)
+3. Image quality (compare with `cache_backend=None`)
+
+---
+
+## Troubleshooting
+
+### Issue: Cache not applied
+
+**Symptoms:** No speedup observed, no cache-related log messages.
+
+**Causes & Solutions:**
+
+- **Enabler not registered:**
+
+**Problem:** Pipeline name not in `CUSTOM_DIT_ENABLERS` registry.
+
+**Solution:** Verify `pipeline.__class__.__name__` matches the registry key and add your enabler to `CUSTOM_DIT_ENABLERS`.
+
+### Issue: Quality degradation
+
+**Symptoms:** Generated images have artifacts or lower quality compared to non-cached inference.
+
+**Causes & Solutions:**
+
+- **Cache parameters too aggressive:**
+
+**Solution:**
+```python
+cache_config={
+    "residual_diff_threshold": 0.12,  # Lower from 0.24 (try 0.12-0.18)
+    "max_warmup_steps": 6,            # Increase from 4 (try 6-8)
+    "max_continuous_cached_steps": 2, # Reduce if higher
+}
+```
+
+Check the [user guide for cache_dit](../../user_guide/diffusion/cache_dit_acceleration.md) for more adjustable parameters.
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Model | Path | Pattern | Notes |
+|-------|------|---------|-------|
+| **Standard DiT** | [`cache_dit_backend.py::enable_cache_for_dit`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/cache_dit_backend/#vllm_omni.diffusion.cache.cache_dit_backend.enable_cache_for_dit) | Default enabler | Single transformer, automatic |
+| **Wan2.2** | [`cache_dit_backend.py::enable_cache_for_wan22`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/cache_dit_backend/#vllm_omni.diffusion.cache.cache_dit_backend.enable_cache_for_wan22) | Dual-transformer | Separate high/low noise transformers |
+| **LongCat** | [`cache_dit_backend.py::enable_cache_for_longcat_image`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/cache_dit_backend/#vllm_omni.diffusion.cache.cache_dit_backend.enable_cache_for_longcat_image) | Multi-block-list | Two block lists in one transformer |
+| **BAGEL** | [`cache_dit_backend.py::enable_cache_for_bagel`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/cache_dit_backend/#vllm_omni.diffusion.cache.cache_dit_backend.enable_cache_for_bagel) | Omni model | Complex architecture |
+
+---
+
+## Summary
+
+Adding cache-dit support:
+
+1. ✅ **Check model type** - Standard models work automatically, custom architectures need enablers
+2. ✅ **Write enabler** (if needed) - Use `BlockAdapter` for complex architectures
+3. ✅ **Register enabler** (if needed) - Add to `CUSTOM_DIT_ENABLERS` dictionary
+4. ✅ **Return refresh function** (if needed) - Handle `num_inference_steps` changes
+5. ✅ **Test** - Verify with `cache_backend="cache_dit"`
+
+For most models, the default enabler is sufficient. Only write custom enablers for complex architectures!

docs/design/feature/cfg_parallel.md

@@ -0,0 +1,249 @@
+# CFG-Parallel
+
+This section describes how to add CFG-Parallel (Classifier-Free Guidance Parallel) to a diffusion pipeline. We use the Qwen-Image pipeline as the reference implementation.
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Step-by-Step Implementation](#step-by-step-implementation)
+- [Customization](#customization)
+- [Testing](#testing)
+- [Troubleshooting](#troubleshooting)
+- [Reference Implementations](#reference-implementations)
+- [Summary](#summary)
+
+---
+
+## Overview
+
+### What is CFG-Parallel?
+
+In standard Classifier-Free Guidance, each diffusion step requires two forward passes through the transformer:
+
+1. **Positive/Conditional**: Guided by the text prompt
+2. **Negative/Unconditional**: Typically using empty or negative prompt
+
+CFG-Parallel eliminates this bottleneck by distributing the two forward passes across different GPU ranks, allowing them to execute simultaneously rather than sequentially.
+
+### Architecture
+
+vLLM-omni provides `CFGParallelMixin` that encapsulates all CFG parallel logic. Pipelines inherit from this mixin and implement a `diffuse()` method that orchestrates the denoising loop.
+
+| Method | Purpose | Automatic Behavior |
+|--------|---------|-------------------|
+| [`predict_noise_maybe_with_cfg()`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/distributed/cfg_parallel/) | Predict noise with CFG | Detects parallel mode, distributes computation, gathers results |
+| [`scheduler_step_maybe_with_cfg()`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/distributed/cfg_parallel/) | Step scheduler with sync | Rank 0 steps, broadcasts latents to all ranks |
+| [`combine_cfg_noise()`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/distributed/cfg_parallel/) | Combine positive/negative | Applies CFG formula with optional normalization |
+| [`predict_noise()`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/distributed/cfg_parallel/) | Forward pass wrapper | Override for custom transformer calls |
+| [`cfg_normalize_function()`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/distributed/cfg_parallel/) | Normalize CFG output | Override for custom normalization |
+
+### How It Works
+
+`predict_noise_maybe_with_cfg()` automatically detects and switches between two execution modes:
+
+- **CFG-Parallel mode** (when `cfg_world_size > 1`):
+  - Rank 0 computes positive prompt prediction
+  - Rank 1 computes negative prompt prediction
+  - Results are gathered via `all_gather()`
+  - Combined on rank 0 using CFG formula
+
+- **Sequential mode** (when `cfg_world_size == 1`):
+  - Single rank computes both positive and negative predictions
+  - Directly combines them with CFG formula
+
+`scheduler_step_maybe_with_cfg()` ensures consistent latent states across all ranks:
+
+- **CFG-Parallel mode**:
+  - Only rank 0 performs the scheduler step (applies noise prediction to update latents)
+  - Updated latents are broadcast to all other ranks via `broadcast()`
+  - All ranks maintain synchronized latent states for the next iteration
+
+- **Sequential mode**:
+  - Single rank directly performs the scheduler step
+  - No synchronization needed
+
+---
+
+## Step-by-Step Implementation
+
+### Step 1: Inherit `CFGParallelMixin`
+
+Allow your pipeline to inherit from `CFGParallelMixin` and implements the `diffuse()` method for your specific model.
+
+**Example (Qwen-Image):**
+
+```python
+from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
+import torch.nn as nn
+class YourModelPipeline(nn.Module, CFGParallelMixin):
+    def diffuse(self, ...) -> torch.Tensor:
+        for i, t in enumerate(timesteps):
+            # Prepare positive_kwargs (conditional) and negative_kwargs (unconditional)
+            positive_kwargs = {...}  # hidden_states, encoder_hidden_states, etc.
+            negative_kwargs = {...} if do_true_cfg else None
+
+            # Key method 1: Predict noise with automatic CFG parallel handling
+            noise_pred = self.predict_noise_maybe_with_cfg(
+                do_true_cfg=do_true_cfg,
+                true_cfg_scale=true_cfg_scale,
+                positive_kwargs=positive_kwargs,
+                negative_kwargs=negative_kwargs,
+            )
+
+            # Key method 2: Step scheduler with automatic CFG synchronization
+            latents = self.scheduler_step_maybe_with_cfg(
+                noise_pred, t, latents, do_true_cfg
+            )
+
+        return latents
+```
+
+**Key Points:**
+
+- `positive_kwargs`: transformer arguments for conditional (text-guided) prediction
+- `negative_kwargs`: transformer arguments for unconditional prediction (set to `None` if CFG disabled)
+- For image editing pipelines, add `output_slice=image_seq_len` to extract the generative image portion
+
+### Step 2: Call `diffuse`
+
+Call `self.diffuse` in your pipeline's forward function:
+
+```python
+import torch.nn as nn
+class YourModelPipeline(nn.Module, CFGParallelMixin):
+    def forward(
+        self,
+        prompt: str,
+        negative_prompt: str | None = None,
+        guidance_scale: float = 3.5,
+        num_inference_steps: int = 50,
+        **kwargs,
+    ):
+        # Encode prompts, Initialize latents, Get timesteps
+        ...
+        # Run diffusion loop (calls the mixin's diffuse method)
+        latents = self.diffuse(
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            negative_prompt_embeds=negative_embeds,
+            negative_prompt_embeds_mask=negative_mask,
+            latents=latents,
+            timesteps=timesteps,
+            do_true_cfg=do_true_cfg,
+            true_cfg_scale=guidance_scale,
+            ...
+        )
+```
+
+---
+
+## Customization
+
+### Override `predict_noise()` for Custom Transformer Calls
+
+If your transformer requires custom prediction function, you can rewrite `predict_noise` function. Taking Wan2.2 as an example, which has two transformer models. The actual transformer to be called is determined by `self.transformer`.
+
+```python
+class Wan22Pipeline(nn.Module, CFGParallelMixin):
+    def predict_noise(self, current_model: nn.Module | None = None, **kwargs: Any) -> torch.Tensor:
+        if current_model is None:
+            current_model = self.transformer
+        return current_model(**kwargs)[0]
+```
+
+
+### Override `cfg_normalize_function()` for Custom Normalization
+
+Some models have their own normalization function. Taking LongCat Image model as an example:
+
+```python
+class LongCatImagePipeline(nn.Module, CFGParallelMixin):
+    def cfg_normalize_function(self, noise_pred, comb_pred, cfg_renorm_min=0.0):
+        """
+        Normalize the combined noise prediction.
+        """
+        cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+        noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+        scale = (cond_norm / (noise_norm + 1e-8)).clamp(min=cfg_renorm_min, max=1.0)
+        noise_pred = comb_pred * scale
+        return noise_pred
+
+        # The original cfg_normalize_function function in CFGParallelMixin
+        # cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+        # noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+        # noise_pred = comb_pred * (cond_norm / noise_norm)
+        # return noise_pred
+```
+
+---
+
+## Testing
+
+After adding CFG-Parallel support, test with:
+
+```bash
+cd examples/offline_inference/text_to_image
+python text_to_image.py \
+    --model Your-org/your-model \
+    --prompt "a cup of coffee on the table" \
+    --negative-prompt "ugly, unclear" \
+    --cfg-scale 4.0 \
+    --num-inference-steps 50 \
+    --output "cfg_enabled.png" \
+    --cfg-parallel-size 2
+```
+
+**Verify:**
+
+1. Check logs for CFG parallel being activated
+2. Record the `e2e_time_ms` in the log and compare with CFG-Parallel disabled
+3. Compare the generated result quality with baseline
+4. Record comparison results in your PR
+
+---
+
+## Troubleshooting
+
+### Issue: CFG parallel not activating
+
+**Symptoms:** Generation still slow, logs don't show CFG parallel being used.
+
+**Causes & Solutions:**
+
+- **CFG is not enabled:**
+
+**Problem:** Guidance scale too low or negative prompt not provided.
+
+**Solution:** Ensure `guidance_scale > 1.0` and negative prompt is provided:
+```python
+images = pipeline(
+    prompt="a cat",
+    negative_prompt="",  # Must provide (even if empty)
+    guidance_scale=3.5,   # Must be > 1.0
+)
+```
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Model | Path | Pattern | Notes |
+|-------|------|---------|-------|
+| **Qwen-Image** | `vllm_omni/diffusion/models/qwen_image/cfg_parallel.py` | Mixin | Dual-stream transformer |
+| **Qwen-Image-Edit** | `vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_edit.py` | Mixin | Image editing with `output_slice` |
+| **Wan2.2** | `vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py` | Mixin | Dual-transformer architecture |
+| **CFGParallelMixin** | `vllm_omni/diffusion/distributed/cfg_parallel.py` | Base implementation | Core mixin class |
+
+---
+
+## Summary
+
+Adding CFG-Parallel support:
+
+1. ✅ **Create mixin** - Inherit from `CFGParallelMixin` and implement `diffuse()` method
+2. ✅ **(Optional) Customize** - Override `predict_noise()` or `cfg_normalize_function()` for custom behavior
+3. ✅ **Test** - Verify with `--cfg-parallel-size 2` and compare performance

docs/design/feature/disaggregated_inference.md

@@ -6,7 +6,8 @@ This guide explains how to configure and use distributed connectors
 Backend-specific setup lives in separate docs:
 
 - [SharedMemoryConnector](omni_connectors/shared_memory_connector.md)
-- [MooncakeConnector](omni_connectors/mooncake_connector.md)
+- [MooncakeStoreConnector](omni_connectors/mooncake_store_connector.md)
+- [MooncakeTransferEngineConnector](omni_connectors/mooncake_transfer_engine_connector.md)
 - [YuanrongConnector](omni_connectors/yuanrong_connector.md)
 
 ## Overview
@@ -19,7 +20,8 @@ Current connectors operate in D2H2D (device to host to device) mode.
 | Use Case | Recommended Connector | Notes |
 | :--- | :--- | :--- |
 | Single node | SharedMemoryConnector | Auto-configured if no connector is specified. |
-| Multi node (Mooncake) | MooncakeConnector | Requires Mooncake Master + metadata server. |
+| Multi node (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. |
 
 ## Core API

docs/design/feature/hsdp.md

@@ -0,0 +1,146 @@
+# HSDP
+
+This section describes how to add HSDP (Hybrid Sharded Data Parallel) support to a diffusion transformer model. We use the Wan2.2 transformer as the reference implementation.
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Step-by-Step Implementation](#step-by-step-implementation)
+- [Testing](#testing)
+- [Reference Implementations](#reference-implementations)
+
+---
+
+## Overview
+
+### What is HSDP?
+
+HSDP (Hybrid Sharded Data Parallel) is a memory optimization technique that **shards model weights** across multiple GPUs using PyTorch's FSDP2. Unlike Tensor Parallelism which splits computation, HSDP:
+
+- Shards weights across GPUs to reduce per-GPU memory usage
+- Gathers weights on-demand during forward passes
+- Can work standalone or combined with other parallelism (e.g., Sequence Parallel)
+
+This enables inference of large models (e.g., Wan2.2 14B) on GPUs with limited memory.
+
+**Important constraints:**
+- HSDP cannot be used with Tensor Parallelism
+- For standalone HSDP (no other parallelism), `hsdp_shard_size` must be specified explicitly
+
+### Architecture
+
+HSDP implementation relies on:
+
+1. **`_hsdp_shard_conditions`**: Model attribute specifying which modules to shard
+2. **`apply_hsdp_to_model`**: Function that applies FSDP2 sharding
+3. **`HSDPInferenceConfig`**: Runtime configuration for HSDP
+
+---
+
+## Step-by-Step Implementation
+
+### Step 1: Identify Modules to Shard
+
+Determine which modules in your transformer should be sharded. Typically, these are:
+
+- Transformer blocks (e.g., `blocks.0`, `blocks.1`, ...)
+- Large submodules with significant weight memory
+
+**Key questions:**
+- Which modules have the largest weights?
+- Which modules are repeated (like transformer blocks)?
+
+### Step 2: Define Shard Conditions
+
+Add `_hsdp_shard_conditions` to your model class. This is a list of functions that return `True` for modules that should be sharded.
+
+**Example (Transformer Blocks):**
+
+```python
+class MyTransformerModel(nn.Module):
+
+    @staticmethod
+    def _is_transformer_block(name: str, module) -> bool:
+        """Match transformer blocks for HSDP sharding.
+
+        Args:
+            name: Module name from named_modules() (e.g., "blocks.0", "blocks.0.attn")
+            module: The module instance
+
+        Returns:
+            True if this module should be sharded
+        """
+        return "blocks" in name and name.split(".")[-1].isdigit()
+
+    _hsdp_shard_conditions = [_is_transformer_block]
+```
+
+**Multiple Conditions Example:**
+
+```python
+class MyModel(nn.Module):
+
+    @staticmethod
+    def _is_transformer_block(name: str, module) -> bool:
+        return "blocks" in name and name.split(".")[-1].isdigit()
+
+    @staticmethod
+    def _is_moe_expert(name: str, module) -> bool:
+        # Also shard MoE expert layers
+        return "experts" in name and name.split(".")[-1].isdigit()
+
+    # Module is sharded if ANY condition returns True
+    _hsdp_shard_conditions = [_is_transformer_block, _is_moe_expert]
+```
+
+---
+
+## Testing
+
+After adding HSDP support, test with:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+parallel_config = DiffusionParallelConfig(
+    use_hsdp=True,
+    hsdp_shard_size=8,  # Shard across 8 GPUs
+)
+omni = Omni(model="your-model-name", parallel_config=parallel_config)
+
+output = omni.generate(
+    "a cup of coffee on the table",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+**Or via command line:**
+
+```bash
+vllm serve Your-org/your-model --omni --port 8091 --use-hsdp
+```
+
+**Verify:**
+
+1. Check logs for "HSDP Inference: replicate_size=..., shard_size=..."
+2. Check logs for "Sharded N modules + root"
+3. Verify memory usage is reduced proportionally
+4. Compare generated output quality with HSDP disabled
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Model | Path | Notes |
+|-------|------|-------|
+| **Wan2.2** | `vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py` | Reference implementation |
+| **HSDP Core** | `vllm_omni/diffusion/distributed/hsdp.py` | `apply_hsdp_to_model`, `shard_model` |
+| **HSDP Tests** | `tests/diffusion/distributed/test_hsdp.py` | Unit tests |
+
+---

docs/design/feature/omni_connectors/mooncake_store_connector.md

@@ -1,4 +1,4 @@
-# MooncakeConnector
+# MooncakeStoreConnector
 
 ## When to Use
 
@@ -38,7 +38,7 @@ Define the connector in runtime:
 runtime:
   connectors:
     connector_of_mooncake:
-      name: MooncakeConnector
+      name: MooncakeStoreConnector
       extra:
         host: "127.0.0.1"
         metadata_server: "http://<MASTER_IP>:8080/metadata"

docs/design/feature/omni_connectors/mooncake_transfer_engine_connector.md

@@ -0,0 +1,207 @@
+# MooncakeTransferEngineConnector
+
+## When to Use
+
+Best for high-performance multi-node data transfer between stages using Mooncake
+Transfer Engine. Supports both RDMA and TCP protocols with a managed memory pool,
+zero-copy deserialization, and optional GPUDirect RDMA. Applicable to any
+inter-stage data (KV caches, request payloads, etc.), not limited to KV cache transfer.
+
+Compared to `MooncakeStoreConnector` (TCP key-value store), this connector
+provides **~60x faster** data transfer via RDMA direct memory access.
+
+## Installation
+
+```bash
+pip install mooncake-transfer-engine
+```
+
+Ensure RDMA drivers are installed on all nodes (e.g., Mellanox OFED for
+InfiniBand/RoCE NICs).
+
+## Configuration
+
+Define the connector in runtime:
+
+```yaml
+runtime:
+  connectors:
+    rdma_connector:
+      name: MooncakeTransferEngineConnector
+      extra:
+        host: "auto"                  # Auto-detect local RDMA IP
+        zmq_port: 50051               # ZMQ base port (see "Port Offset Scheme" below)
+        protocol: "rdma"              # "rdma" or "tcp"
+        device_name: ""               # RDMA device (e.g., "mlx5_0"), empty for auto-detect
+        memory_pool_size: 2147483648  # 2GB memory pool
+        memory_pool_device: "cpu"     # "cpu" for pinned memory, "cuda" for GPUDirect RDMA
+```
+
+Wire stages to the connector:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    output_connectors:
+      to_stage_1: rdma_connector
+
+  - stage_id: 1
+    input_connectors:
+      from_stage_0: rdma_connector
+```
+
+## Parameters
+
+### Required
+
+| Parameter | Description |
+|---|---|
+| `role` | **Internal, do not set manually.** Auto-injected by the orchestration layer (`"sender"` for `output_connectors`, `"receiver"` for `input_connectors`). Defaults to `"sender"` if omitted. |
+| `host` | Local IP address for RDMA. `"auto"` detects from network interfaces. |
+| `protocol` | Transport protocol: `"rdma"` (InfiniBand/RoCE) or `"tcp"`. |
+
+### Memory Pool
+
+| Parameter | Default | Description |
+|---|---|---|
+| `memory_pool_size` | 1 GB | Total size of the RDMA-registered memory pool in bytes. |
+| `memory_pool_device` | `"cpu"` | `"cpu"`: pinned host memory (recommended). `"cuda"`: GPU VRAM for GPUDirect RDMA (requires NIC-GPU direct PCIe connectivity). |
+
+### Networking
+
+| Parameter | Default | Description |
+|---|---|---|
+| `zmq_port` | 50051 | ZMQ **base** port. The orchestration layer computes the actual port as `base + purpose_offset + stage_offset` (see table below). Users only set this base value. |
+| `sender_host` | `None` | **Internal.** Receiver-side only — dynamically resolved via `update_sender_info()`. Not needed in YAML. |
+| `sender_zmq_port` | `None` | **Internal.** Receiver-side only — defaults to the sender's adjusted port. Not needed in YAML. |
+| `device_name` | `""` | RDMA device name (e.g., `"mlx5_0"`). Empty for auto-detect. Can also be set via `RDMA_DEVICE_NAME` env var. |
+
+#### ZMQ Port Offset Scheme
+
+To avoid port conflicts when multiple edges, purposes, DP replicas, or TP ranks share the same node, the actual ZMQ port is computed as:
+
+```
+side_channel_port = zmq_port + purpose_offset + stage_offset + dp_index * tp_size
+sender_listen     = side_channel_port + tp_rank
+receiver_connect  = remote_side_channel_port + tp_rank
+```
+
+| Component | Value | Description |
+|---|---|---|
+| `zmq_port` | 50051 (default) | Base port from YAML config |
+| `purpose_offset` | `request_forwarding` = 0, `kv_transfer` = 100 | Separates control-plane vs KV-cache connections |
+| `stage_offset` | `int(from_stage)` (0, 1, 2...) | Separates edges from different source stages |
+| `dp_index * tp_size` | e.g., DP1 × TP2 = 2 | Each DP replica reserves a port range of size `tp_size` (following vLLM convention: `VLLM_MOONCAKE_BOOTSTRAP_PORT + dp_index * tp_size`) |
+| `tp_rank` | 0, 1, 2... | Each TP rank within a DP replica uses its own port |
+| orchestrator | +200 | Extra offset when caller is the orchestrator (avoids collision with stage workers on the same node) |
+
+**Example** (base=50051, stage 0→1, DP=2, TP=2, kv_transfer):
+
+| Caller | DP | TP rank | Port |
+|---|---|---|---|
+| Stage worker | DP0 | rank 0 | `50051 + 100 + 0 + 0×2 + 0 = 50151` |
+| Stage worker | DP0 | rank 1 | `50051 + 100 + 0 + 0×2 + 1 = 50152` |
+| Stage worker | DP1 | rank 0 | `50051 + 100 + 0 + 1×2 + 0 = 50153` |
+| Stage worker | DP1 | rank 1 | `50051 + 100 + 0 + 1×2 + 1 = 50154` |
+| Orchestrator | — | — | `50051 + 200 + 0 = 50251` |
+
+## Memory Pool Modes
+
+| Mode | Config | Data Flow | Best For |
+|---|---|---|---|
+| CPU Pinned | `memory_pool_device: "cpu"` | GPU → CPU pool → RDMA → CPU pool → GPU | Most hardware topologies (recommended) |
+| GPUDirect | `memory_pool_device: "cuda"` | GPU → GPU pool → RDMA (NIC reads GPU BAR1) → GPU pool | NIC-GPU direct PCIe (PIX topology) |
+
+> **Note**: GPUDirect RDMA requires the NIC and GPU to share a direct PCIe
+> switch (PIX topology). On systems where they are connected via PXB or NODE,
+> CPU pinned memory is faster due to GPU BAR1 bandwidth limitations.
+
+## Environment Variables
+
+| Variable | Description |
+|---|---|
+| `RDMA_DEVICE_NAME` | Override RDMA device name (e.g., `mlx5_0`). |
+| `MC_IB_PCI_RELAXED_ORDERING` | Set to `1` to enable PCIe relaxed ordering for GPUDirect. |
+
+## Docker / Container Setup
+
+RDMA requires host-level device access:
+
+```bash
+docker run -it \
+    --cap-add=SYS_PTRACE \
+    --cap-add=IPC_LOCK \
+    --security-opt seccomp=unconfined \
+    --network=host \
+    --device=/dev/infiniband \
+    -v /sys/class/infiniband:/sys/class/infiniband:ro \
+    your-image:tag
+```
+
+## Performance
+
+Benchmark results on H800 GPUs with mlx5_0 RDMA NIC (~186 MB KV cache):
+
+| Metric | MooncakeStoreConnector | MooncakeTransferEngineConnector (CPU) |
+|---|---|---|
+| KV transfer wall time | ~810 ms | **~14 ms** |
+| RDMA throughput | N/A (TCP) | ~22 GB/s |
+| Speedup | 1x | **58x** |
+
+## Troubleshooting
+
+### Quick Diagnostics
+
+```bash
+# 1. Check RDMA devices and link status
+ibdev2netdev
+# Expected: "mlx5_X port 1 ==> <iface> (Up)"
+# RoCE devices map to Ethernet interfaces (e.g., enp75s0f0)
+# IB devices map to ib0, ib1, etc.
+
+# 2. Check InfiniBand device details
+ibstat
+
+# 3. Verify /dev/infiniband is accessible (required in containers)
+ls /dev/infiniband/
+
+# 4. Check Mooncake installation
+python -c "from mooncake.engine import TransferEngine; print('OK')"
+
+# 5. Check environment variables
+echo "RDMA_DEVICE_NAME=${RDMA_DEVICE_NAME:-<not set>}"
+echo "MC_IB_PCI_RELAXED_ORDERING=${MC_IB_PCI_RELAXED_ORDERING:-<not set>}"
+```
+
+### Common Issues
+
+| Symptom | Cause | Fix |
+|---|---|---|
+| `Failed to modify QP to RTR` | Cross-NIC QP handshake failure (multi-NIC DGX) | Set `device_name` to a single RoCE NIC (e.g., `mlx5_2`) or set `RDMA_DEVICE_NAME` env var |
+| `transport retry counter exceeded` | RDMA path between incompatible NICs | Same as above — restrict to one NIC |
+| `zmq.error.Again: Resource temporarily unavailable` | ZMQ recv timeout (transfer took too long) | Check NIC selection; increase data may need longer timeout |
+| `Mooncake Engine initialization failed` | Missing RDMA drivers or `/dev/infiniband` | Install Mellanox OFED; in Docker add `--device=/dev/infiniband` |
+| `MemoryError` in allocator | Memory pool too small for payload | Increase `memory_pool_size` |
+| GPU transfer slower than CPU | GPU BAR1 bandwidth limitation (PXB/NODE topology) | Use `memory_pool_device: "cpu"` instead of `"cuda"` |
+
+### Multi-NIC Environments (DGX)
+
+On DGX machines with 12+ RDMA NICs, only RoCE NICs (with a bound network
+interface) reliably support loopback. IB-only NICs may fail cross-NIC QP
+handshakes. To identify RoCE NICs:
+
+```bash
+ibdev2netdev | grep -v "ib[0-9]"
+# RoCE devices show Ethernet interface names like enp75s0f0
+```
+
+Then configure the connector:
+```yaml
+device_name: "mlx5_2"  # or set RDMA_DEVICE_NAME=mlx5_2
+```
+
+See the RDMA Test README in tests/distributed/omni_connectors/README.md
+for test-specific setup instructions.
+
+For more details on the underlying engine, refer to the
+[Mooncake repository](https://github.com/kvcache-ai/Mooncake).

docs/design/feature/ray_based_execution.md

@@ -45,7 +45,7 @@ ray start --address=<HEAD_NODE_IP>:6399
 
 When running on Ray, the system automatically adapts its communication strategy:
 
-*   **Cross-Node**: Recommended to use `MooncakeConnector` (requires separate configuration).
+*   **Cross-Node**: Recommended to use `MooncakeTransferEngineConnector` (RDMA, fastest) or `MooncakeStoreConnector` (TCP fallback).
 *   **Same-Node**: Can still use `SharedMemoryConnector` for efficiency, or Ray's native object store (plasma).
 *   **SHM threshold default differs**: when `worker_backend="ray"`, the SharedMemoryConnector default threshold is set to `sys.maxsize`, which forces payloads to go inline (no SHM). Override `shm_threshold_bytes` in the connector config if you want SHM for Ray runs.
 

docs/design/feature/sequence_parallel.md

@@ -0,0 +1,456 @@
+# Sequence Parallel
+
+This section describes how to add Sequence Parallel (SP) to a diffusion transformer model. We use the Qwen-Image transformer and Wan2.2 transformer as reference implementations.
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Approach 1: Non-Intrusive `_sp_plan` (Recommended)](#approach-1-non-intrusive-_sp_plan-recommended)
+- [Approach 2: Intrusive Modification (For Complex Cases)](#approach-2-intrusive-modification-for-complex-cases)
+- [Testing](#testing)
+- [Troubleshooting](#troubleshooting)
+- [Reference Implementations](#reference-implementations)
+- [Summary](#summary)
+
+---
+
+## Overview
+
+
+### What is Sequence Parallel?
+
+**Terminology Note:** Our "Sequence Parallelism" (SP) corresponds to "Context Parallelism" (CP) in the [diffusers library](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/_modeling_parallel.py). We use "Sequence Parallelism" to align with vLLM-Omni's terminology.
+
+Diffusion transformers process long sequences of image patches or video frames. For high-resolution generation, these sequences can become very large. Enabling SP allows each GPU to process only a portion of the sequence, with attention mechanisms (Ulysses/Ring) handling cross-GPU communication transparently.
+
+### Architecture
+
+The major APIs for Sequence Parallel:
+
+```python
+from vllm_omni.diffusion.distributed.sp_plan import (
+    SequenceParallelInput,   # For sharding (splitting) tensors
+    SequenceParallelOutput,  # For gathering tensors
+)
+from vllm_omni.diffusion.distributed.sp_sharding import sp_shard, sp_gather
+```
+
+| Method/Class | Purpose | Behavior |
+|--------------|---------|----------|
+| `SequenceParallelInput` | Declare input sharding in `_sp_plan` | Auto-shards tensors at module input |
+| `SequenceParallelOutput` | Declare output gathering in `_sp_plan` | Auto-gathers tensors at module output |
+| `sp_shard()` | Manual tensor sharding | Splits tensor across SP workers |
+| `sp_gather()` | Manual tensor gathering | Gathers sharded tensors from all workers |
+
+---
+
+## Approach 1: Non-Intrusive `_sp_plan` (Recommended)
+
+The `_sp_plan` mechanism allows SP **without modifying `forward()` logic**. The framework automatically registers hooks to shard inputs and gather outputs at module boundaries.
+
+**When to use:**
+- Standard transformer architectures
+- Tensor operations happen at `nn.Module` boundaries
+- Predictable sharding/gathering patterns
+
+**How it works:**
+1. Declare `_sp_plan` dict in your transformer class
+2. Framework automatically applies hooks when `sequence_parallel_size > 1`
+3. Hooks shard/gather tensors at specified module boundaries
+4. Attention layers handle cross-GPU communication internally
+
+```python
+class StandardTransformer(nn.Module):
+    _sp_plan = {
+        # Shard hidden_states at first transformer block input
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        # Gather at final output projection
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+`StandardTransformer` has a transformer blocks list `self.blocks = nn.ModuleList([...])`, and a projection output layer `self.proj_out`. The `_sp_plan` above defines that when SP is enabled, sharding the input tensor to the first transformer block, and gathering the sharded tensor at the final output projection layer.
+
+**Requirements:**
+- Tensor operations that need sharding/gathering must happen at **`nn.Module` boundaries**
+- Inline Python operations (e.g., `torch.cat`, `pad_sequence`) **cannot be hooked**
+
+**Solution for inline operations:** Extract into a submodule (see Step 2 below).
+
+### Step 1: Understand Module Boundaries
+
+Identify where tensors need to be sharded or gathered in your model's forward pass:
+
+```python
+class MyTransformer(nn.Module):
+    def __init__(self):
+        self.patch_embed = PatchEmbed()      # ← Boundary 1
+        self.pos_embed = RoPE()              # ← Boundary 2
+        self.blocks = nn.ModuleList([...])   # ← Boundary 3
+        self.norm_out = LayerNorm()
+        self.proj_out = Linear()             # ← Boundary 4
+
+    def forward(self, x):
+        x = self.patch_embed(x)              # ← Shard before this?
+        pos = self.pos_embed(x)              # ← Shard RoPE outputs?
+        for block in self.blocks:
+            x = block(x, pos)                # ← Blocks process sharded x
+        x = self.norm_out(x)
+        output = self.proj_out(x)            # ← Gather after this?
+        return output
+```
+
+### Step 2: Handle Inline Operations
+
+If your `forward()` contains inline tensor operations, **extract them into submodules**.
+
+**Example: Z-Image concatenates image + text features inline**
+
+```python
+# ❌ BAD: Inline operation - hooks cannot intercept
+class ZImageTransformer(nn.Module):
+    def forward(self, x, cap_feats):
+        # This concatenation happens inline - _sp_plan can't shard it!
+        unified = torch.cat([x, cap_feats], dim=1)
+
+        for layer in self.layers:
+            unified = layer(unified)
+
+        return unified
+
+# ✅ GOOD: Extract into submodule
+class UnifiedPrepare(nn.Module):
+    """Submodule to concatenate image and text features."""
+    def forward(self, x, cap_feats):
+        return torch.cat([x, cap_feats], dim=1)
+
+class ZImageTransformer(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.unified_prepare = UnifiedPrepare()  # Now a module!
+        self.layers = nn.ModuleList([...])
+
+    def forward(self, x, cap_feats):
+        # Now _sp_plan can shard the output of unified_prepare!
+        unified = self.unified_prepare(x, cap_feats)
+
+        for layer in self.layers:
+            unified = layer(unified)
+
+        return unified
+```
+
+**Other common cases:**
+- `pad_sequence()` → `PadSequenceModule`
+- `torch.cat()` → `ConcatModule`
+- `tensor.reshape()` → `ReshapeModule`
+- Complex preprocessing → `PreprocessModule`
+
+### Step 3: Write `_sp_plan` for Your Model
+
+Create a class-level `_sp_plan` dictionary specifying where to shard/gather tensors.
+
+Typically, there are two patterns for diffusion models:
+
+**Pattern 1: Shard at first block, gather at output projection**
+
+Most common pattern for standard transformers:
+
+```python
+from vllm_omni.diffusion.distributed.sp_plan import (
+    SequenceParallelInput,   # For sharding (splitting) tensors
+    SequenceParallelOutput,  # For gathering tensors
+)
+class StandardTransformer(nn.Module):
+    _sp_plan = {
+        # Shard hidden_states at first transformer block input
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        # Gather at final output projection
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+**Pattern 2: Shard RoPE embeddings separately**
+
+When RoPE is computed in a separate module:
+
+```python
+from vllm_omni.diffusion.distributed.sp_plan import (
+    SequenceParallelInput,   # For sharding (splitting) tensors
+    SequenceParallelOutput,  # For gathering tensors
+)
+class TransformerWithRoPE(nn.Module):
+    _sp_plan = {
+        # Shard RoPE module OUTPUTS (returns tuple of cos, sin)
+        "rope": {
+            0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),  # cos
+            1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),  # sin
+        },
+        # Shard transformer block INPUT
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        # Gather at output
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+### API Reference
+
+**SequenceParallelInput Parameters:**
+
+| Parameter | Type | Description |
+|-----------|------|-------------|
+| `split_dim` | int | Dimension to split (usually `1` for sequence) |
+| `expected_dims` | int \| None | Expected tensor rank for validation (optional) |
+| `split_output` | bool | `False`: shard **input** params; `True`: shard **output** tensors |
+| `auto_pad` | bool | Auto-pad if sequence not divisible by world_size (default: `False`) |
+
+**SequenceParallelOutput Parameters:**
+
+| Parameter | Type | Description |
+|-----------|------|-------------|
+| `gather_dim` | int | Dimension to gather (usually `1` for sequence) |
+| `expected_dims` | int \| None | Expected tensor rank for validation (optional) |
+
+**Module Naming Conventions:**
+
+| Key | Meaning | Python equivalent |
+|-----|---------|-------------------|
+| `""` | Root model | `model` |
+| `"blocks.0"` | First element of ModuleList | `model.blocks[0]` |
+| `"blocks.*"` | All elements of ModuleList | `for b in model.blocks` |
+| `"rope"` | Named submodule | `model.rope` |
+| `"outputs.main"` | ModuleDict entry | `model.outputs["main"]` |
+
+**Dictionary Value Types:**
+
+| Key type | `split_output` | Description |
+|----------|----------------|-------------|
+| `"param_name"` (str) | `False` | Shard **input parameter** by name |
+| `0`, `1`, ... (int) | `True` | Shard **output tuple** by index |
+
+---
+
+## Approach 2: Intrusive Modification (For Complex Cases)
+
+For models with dynamic sharding logic that cannot be expressed via `_sp_plan`, manually insert shard/gather calls.
+
+**When to use:**
+- Dynamic/conditional sharding logic
+- Complex tensor manipulations that can't be encapsulated
+- Temporary workaround during development
+
+```python
+from vllm_omni.diffusion.distributed.sp_sharding import sp_shard, sp_gather
+
+def forward(self, hidden_states, ...):
+    if self.parallel_config.sequence_parallel_size > 1:
+        hidden_states = sp_shard(hidden_states, dim=1)
+
+    # ... computation ...
+
+    if self.parallel_config.sequence_parallel_size > 1:
+        output = sp_gather(output, dim=1)
+
+    return output
+```
+
+---
+
+## Testing
+
+After implementing Sequence Parallel support, thoroughly test your implementation to ensure correctness and performance across different configurations.
+
+**Test Different `sp_size`:**
+
+Test your model with various sequence parallel world sizes to verify correctness and identify optimal configurations:
+
+```bash
+cd examples/offline_inference/text_to_image
+python text_to_image.py \
+    --model Your-org/your-model \
+    --prompt "a cup of coffee on the table" \
+    --num-inference-steps 50 \
+    --ulysses-degree 2 \
+    --ring-degree 2 \
+    --output sp_test_image_ulysses=2_ring=2.png
+```
+
+**Verify:**
+
+1. **Correctness:** Output should be identical across all `sp_size` values
+2. **Speed:** Throughput should remain stable or improve (especially for large sequences)
+3. **Logs:** Check for any shape mismatch or communication errors
+
+**Test with Tensor Parallel:**
+
+Sequence Parallel can be combined with other parallelism strategies:
+
+```bash
+cd examples/offline_inference/text_to_image
+python text_to_image.py \
+    --model Your-org/your-model \
+    --prompt "a cup of coffee on the table" \
+    --num-inference-steps 50 \
+    --ulysses-degree 2 \
+    --tensor-parallel-size 2 \
+    --output sp_test_image_ulysses=2_tp=2.png
+```
+
+---
+
+## Troubleshooting
+
+### Issue: Shape mismatch errors
+
+**Symptoms:** `RuntimeError: shape mismatch` during forward pass.
+
+**Causes & Solutions:**
+
+- **RoPE dimension mismatch:**
+
+**Problem:** RoPE embeddings not sharded, but hidden_states is sharded.
+
+**Solution:** Shard RoPE outputs in `_sp_plan`:
+```python
+_sp_plan = {
+    "rope": {
+        0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),
+        1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),
+    },
+    ...
+}
+```
+
+- **Sequence Length not divisible by sp_size:**
+
+**Problem:** `SequenceParallelInput(auto_pad=False)` - auto_pad should be True to enable automatic sequence padding.
+
+**Solution:** In `SequenceParallelInput`, set `auto_pad=True` and add attention mask support.
+
+> **Experimental Feature:** `auto_pad=True` is an experimental feature and may be changed in the future. We plan to improve this solution to involve minimal changes to model files. More details are [here](https://github.com/vllm-project/vllm-omni/issues/1324).
+
+**Constraints of auto_pad:**
+
+| Constraint | Description |
+|------------|-------------|
+| **Attention Backend Compatibility** | The attention backends must support `attention_mask`. Currently only `TORCH_SDPA` and `FLASH_ATTN` (default for diffusion models) are supported. |
+| **Ring Attention Limitation** | Ring attention does not support `attention_mask`. Therefore, when using `auto_pad=True`, the combination of Ulysses + Ring attention is not feasible. |
+
+1. Enable `auto_pad=True` for all sequence-dimension inputs in `_sp_plan`:
+```python
+_sp_plan = {
+    "rope": {
+        0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True, auto_pad=True),
+        1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True, auto_pad=True),
+    },
+    "blocks.0": {
+        "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3, auto_pad=True)
+    },
+    ...
+}
+```
+
+2. Create attention mask dynamically when padding is applied:
+```python
+from vllm_omni.diffusion.forward_context import get_forward_context
+from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata
+
+# In model forward(), before transformer blocks:
+hidden_states_mask = None
+ctx = get_forward_context()
+if ctx.sp_original_seq_len is not None and ctx.sp_padding_size > 0:
+    batch_size = hidden_states.shape[0]
+    padded_seq_len = ctx.sp_original_seq_len + ctx.sp_padding_size
+    hidden_states_mask = torch.ones(batch_size, padded_seq_len, dtype=torch.bool, device=hidden_states.device)
+    hidden_states_mask[:, ctx.sp_original_seq_len:] = False
+
+# Pass mask to attention layers
+attn_metadata = AttentionMetadata(attn_mask=hidden_states_mask) if hidden_states_mask is not None else None
+output = self.attn(query, key, value, attn_metadata)
+```
+
+**Important Quality Considerations:**
+
+While `auto_pad` enables generation for irregular resolutions, be aware of potential quality impacts:
+
+| Aspect | Impact |
+|--------|--------|
+| **Training Distribution** | Models perform best on aspect ratios seen during training (e.g., 1:1, 16:9, 4:3). Unusual ratios like 700x400 (1.75:1) may produce lower quality results. |
+| **Padding Overhead** | Padded positions consume compute even when masked. For best efficiency, prefer resolutions divisible by `sp_size`. |
+
+**Recommendations for users:**
+- Use standard aspect ratios when possible (e.g., 768x432 for 16:9 instead of 700x400)
+- Ensure post-patch dimensions are divisible by `sp_size` for optimal quality
+- Test generation quality when using unusual resolutions
+
+### Issue: Inline operations not sharded
+
+**Symptoms:** Some tensors remain full-sized, not sharded.
+
+**Causes & Solutions:**
+
+- **Operations happen inline in `forward()`, not at module boundaries:**
+
+**Problem:**
+```python
+def forward(self, x, cap):
+    unified = torch.cat([x, cap], dim=1)  # ← Inline operation!
+    # _sp_plan can't hook this
+```
+
+**Solution:** Extract into submodule:
+```python
+class ConcatModule(nn.Module):
+    def forward(self, x, cap):
+        return torch.cat([x, cap], dim=1)
+
+class MyModel(nn.Module):
+    _sp_plan = {
+        "concat": {
+            0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),
+            1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),
+        },
+        ...
+    }
+    def __init__(self):
+        self.concat = ConcatModule()  # Now hookable!
+
+    def forward(self, x, cap):
+        unified = self.concat(x, cap)  # ← Can be sharded via _sp_plan
+```
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Model | Path | Pattern | Notes |
+|-------|------|---------|-------|
+| **Qwen-Image** | `vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py` | Dual-stream + preprocessing | auto_pad, separate RoPE |
+| **Wan2.2** | `vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py` | Dual-Transformer + RoPE | Video transformer |
+| **Z-Image** | `vllm_omni/diffusion/models/z_image/z_image_transformer.py` | Unified sequence | Concatenated input |
+| **SP Plan Types** | `vllm_omni/diffusion/distributed/sp_plan.py` | Type definitions | SequenceParallelInput/Output |
+| **Hook Implementation** | `vllm_omni/diffusion/hooks/sequence_parallel.py` | Hook mechanics | How hooks work |
+| **Tests** | `tests/diffusion/distributed/test_sp_plan_hooks.py` | Test examples | Validation patterns |
+
+---
+
+## Summary
+
+Adding Sequence Parallel support to a transformer:
+
+1. ✅ **Choose approach** - Use `_sp_plan` for standard cases, intrusive modification for complex cases
+2. ✅ **Identify sharding boundaries** - Where should tensors be split/gathered?
+3. ✅ **Extract inline operations** - Move `torch.cat`, `pad_sequence`, etc. to submodules
+4. ✅ **Define `_sp_plan`** - Declare shard/gather points as class attribute
+5. ✅ **Use `auto_pad` for variable lengths** - Support non-uniform sequences
+6. ✅ **Test** - Verify with different `ulysses_degree` and `ring_degree` combinations

docs/design/feature/teacache.md

@@ -0,0 +1,459 @@
+# TeaCache
+
+This section describes how to add TeaCache to a diffusion transformer model. We use the Qwen-Image transformer as the reference implementation.
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Step-by-Step Implementation](#step-by-step-implementation)
+- [Customization](#customization)
+- [Testing](#testing)
+- [Troubleshooting](#troubleshooting)
+- [Reference Implementations](#reference-implementations)
+- [Summary](#summary)
+
+---
+
+## Overview
+
+### What is TeaCache?
+
+TeaCache speeds up diffusion inference by caching transformer block computations when consecutive timesteps are similar. It provides **1.5x-2.0x speedup** with minimal quality loss.
+
+The core insight is that the modulated input (after normalization and timestep conditioning) changes gradually across timesteps. By measuring the L1 distance between consecutive modulated inputs and comparing it to a threshold, TeaCache decides whether to execute the full transformer blocks or reuse the cached residual from the previous step.
+
+vLLM-omni provides a **hook-based** TeaCache system that requires **zero changes to model code**. The hook completely intercepts the transformer's forward pass and implements adaptive caching transparently. This design allows easy integration with any transformer model by simply writing an extractor function.
+
+### Architecture
+
+The TeaCache system consists of three main components:
+
+| Component | Purpose | Location |
+|-----------|---------|----------|
+| [`CacheContext`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/#vllm_omni.diffusion.cache.CacheContext) | Dataclass containing model-specific information for caching | `vllm_omni/diffusion/cache/teacache/context.py` |
+| [`EXTRACTOR_REGISTRY`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/teacache/extractors/#vllm_omni.diffusion.cache.teacache.extractors.EXTRACTOR_REGISTRY) | Maps transformer class names to extractor functions | `vllm_omni/diffusion/cache/teacache/extractors.py` |
+| [`TeaCacheConfig`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/cache/#vllm_omni.diffusion.cache.TeaCacheConfig) | Configuration including thresholds and polynomial coefficients | `vllm_omni/diffusion/cache/teacache/config.py` |
+
+The hook handles all caching logic automatically, including:
+
+- CFG-aware state management (separate states for positive/negative branches)
+- CFG-parallel compatibility
+- L1 distance computation with polynomial rescaling
+- Residual caching and reuse
+
+
+---
+
+## Step-by-Step Implementation
+
+To add TeaCache support for a new model, you need to:
+
+1. Write an **extractor function** that returns a `CacheContext` object
+2. Register the extractor in the `EXTRACTOR_REGISTRY`
+3. Add model-specific polynomial coefficients to `TeaCacheConfig`
+
+### Step 1: Model-Specific Preprocessing
+
+Extract and process model inputs. This typically involves:
+- Embedding image/latent inputs
+- Processing text encoder outputs (if dual-stream)
+- Creating timestep embeddings
+- Computing positional embeddings
+
+**Example (Qwen-Image):**
+
+```python
+def extract_qwen_context(
+    module: nn.Module,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    encoder_hidden_states_mask: torch.Tensor,
+    timestep: torch.Tensor,
+    img_shapes: torch.Tensor,
+    txt_seq_lens: torch.Tensor,
+    guidance: torch.Tensor | None = None,
+    **kwargs: Any,
+) -> CacheContext:
+    # Validate model structure
+    if not hasattr(module, "transformer_blocks") or len(module.transformer_blocks) == 0:
+        raise ValueError("Module must have transformer_blocks")
+
+    # Preprocessing: embed inputs
+    hidden_states = module.img_in(hidden_states)
+    timestep = timestep.to(device=hidden_states.device, dtype=hidden_states.dtype)
+    encoder_hidden_states = module.txt_norm(encoder_hidden_states)
+    encoder_hidden_states = module.txt_in(encoder_hidden_states)
+
+    # Create timestep embedding
+    if guidance is not None:
+        guidance = guidance.to(hidden_states.dtype) * 1000
+    temb = (
+        module.time_text_embed(timestep, hidden_states)
+        if guidance is None
+        else module.time_text_embed(timestep, guidance, hidden_states)
+    )
+
+    # Compute position embeddings
+    image_rotary_emb = module.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)
+```
+
+### Step 2: Extract Modulated Input
+
+The modulated input is used for cache decisions. Extract it from the **first transformer block** after normalization and modulation.
+
+**Example (Qwen-Image):**
+
+```python
+    # Extract modulated input from first transformer block
+    block = module.transformer_blocks[0]
+    img_mod_params = block.img_mod(temb)
+    img_mod1, _ = img_mod_params.chunk(2, dim=-1)
+    img_modulated, _ = block.img_norm1(hidden_states, img_mod1)
+```
+
+**Key Points:**
+
+- Use the **first block** to extract modulated input early
+- Apply the same normalization and modulation as the actual forward pass
+- The tensor should represent the processed features that will change across timesteps
+
+### Step 3: Define Transformer Execution
+
+Create a callable that executes all transformer blocks. This encapsulates the main computation loop.
+
+**Example (Qwen-Image dual-stream):**
+
+```python
+    def run_transformer_blocks():
+        """Execute all Qwen transformer blocks."""
+        h = hidden_states
+        e = encoder_hidden_states
+
+        for block in module.transformer_blocks:
+            e, h = block(
+                hidden_states=h,
+                encoder_hidden_states=e,
+                encoder_hidden_states_mask=encoder_hidden_states_mask,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+            )
+        return (h, e)  # Return both image and text hidden states
+```
+
+**Example (Single-stream model like Flux):**
+
+```python
+    def run_transformer_blocks():
+        """Execute all Flux transformer blocks."""
+        h = hidden_states
+
+        for block in module.transformer_blocks:
+            h = block(h, temb=temb)
+        return (h,)  # Return only image hidden states
+```
+
+**Key Points:**
+
+- Return format:
+- For single-stream models: return `(hidden_states,)`
+- For dual-stream models: return `(hidden_states, encoder_hidden_states)`
+
+### Step 4: Define Postprocessing
+
+Create a callable that applies final transformations to produce the model output.
+
+**Example (Qwen-Image):**
+
+```python
+    return_dict = kwargs.get("return_dict", True)
+
+    def postprocess(h):
+        """Apply Qwen-specific output postprocessing."""
+        h = module.norm_out(h, temb)
+        output = module.proj_out(h)
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
+```
+
+### Step 5: Return CacheContext
+
+Package all information into a `CacheContext` object.
+
+```python
+    return CacheContext(
+        modulated_input=img_modulated,
+        hidden_states=hidden_states,
+        encoder_hidden_states=encoder_hidden_states,  # or None for single-stream
+        temb=temb,
+        run_transformer_blocks=run_transformer_blocks,
+        postprocess=postprocess,
+    )
+```
+
+**CacheContext Fields:**
+
+| Field | Type | Purpose |
+|-------|------|---------|
+| `modulated_input` | `torch.Tensor` | Tensor used for cache decision (similarity comparison) |
+| `hidden_states` | `torch.Tensor` | Current hidden states (will be modified by caching) |
+| `encoder_hidden_states` | `torch.Tensor | None` | Encoder states for dual-stream models, `None` for single-stream |
+| `temb` | `torch.Tensor` | Timestep embedding tensor |
+| `run_transformer_blocks` | `Callable[[], tuple]` | Executes transformer blocks, returns `(hidden_states, [encoder_hidden_states])` |
+| `postprocess` | `Callable[[torch.Tensor], Any]` | Applies final transformations to produce model output |
+| `extra_states` | `dict | None` | Optional dict for additional model-specific state |
+
+### Step 6: Register the Extractor
+
+Add your extractor to the `EXTRACTOR_REGISTRY` in `vllm_omni/diffusion/cache/teacache/extractors.py`:
+
+```python
+EXTRACTOR_REGISTRY: dict[str, Callable] = {
+    "QwenImageTransformer2DModel": extract_qwen_context,
+    "Bagel": extract_bagel_context,
+    "ZImageTransformer2DModel": extract_zimage_context,
+    "YourModelTransformer2DModel": extract_your_model_context,  # Add here
+}
+```
+
+**Key:** Use the transformer class name (`module.__class__.__name__`)
+
+### Step 7: Add Model Coefficients
+
+Add polynomial rescaling coefficients to `vllm_omni/diffusion/cache/teacache/config.py`:
+
+```python
+_MODEL_COEFFICIENTS = {
+    "QwenImageTransformer2DModel": [
+        -4.50000000e02,
+        2.80000000e02,
+        -4.50000000e01,
+        3.20000000e00,
+        -2.00000000e-02,
+    ],
+    "YourModelTransformer2DModel": [  # Add your model's coefficients
+        # 5 polynomial coefficients (can reuse similar model's coefficients initially)
+    ],
+}
+```
+
+
+**Initial approach:** Start with coefficients from a similar model architecture, then tune empirically following [Customization](#customization) section.
+
+---
+
+## Customization
+
+### Coefficient Estimation
+
+While you can start with coefficients from a similar model architecture, estimating custom coefficients for your specific model typically improves TeaCache performance.
+
+**Why Estimate Coefficients?**
+
+The polynomial coefficients rescale L1 distances between consecutive modulated inputs to better predict when cached residuals can be reused. Model-specific coefficients account for:
+
+- Architecture differences (layer count, hidden size, attention patterns)
+- Training data characteristics
+- Noise prediction behavior across timesteps
+
+| Approach | Performance | Effort |
+|----------|-------------|--------|
+| Using defaults from similar model | Within 5-10% of optimal | Low |
+| Estimating custom coefficients | Best performance | Medium |
+
+#### Implement Data Collection Adapter
+
+Add an adapter in `vllm_omni/diffusion/cache/teacache/coefficient_estimator.py`:
+
+```python
+class YourModelAdapter:
+    """Adapter for coefficient estimation on your model."""
+
+    @staticmethod
+    def load_pipeline(model_path: str, device: str, dtype: torch.dtype) -> Any:
+        """Load your diffusion pipeline."""
+        from your_model_package import YourModelPipeline
+
+        pipeline = YourModelPipeline.from_pretrained(
+            model_path,
+            torch_dtype=dtype,
+        )
+        pipeline = pipeline.to(device)
+        return pipeline
+
+    @staticmethod
+    def get_transformer(pipeline: Any) -> tuple[Any, str]:
+        """Extract transformer from pipeline."""
+        return pipeline.transformer, "YourTransformer2DModel"
+
+    @staticmethod
+    def install_hook(transformer: Any, hook: DataCollectionHook) -> None:
+        """Install data collection hook on transformer."""
+        from vllm_omni.diffusion.hooks import HookRegistry
+
+        registry = HookRegistry.get_or_create(transformer)
+        registry.register_hook(hook._HOOK_NAME, hook)
+
+
+# Register your adapter
+_MODEL_ADAPTERS["YourModel"] = YourModelAdapter
+```
+
+#### Collect Data and Estimate
+
+```python
+from vllm_omni.diffusion.cache.teacache.coefficient_estimator import (
+    TeaCacheCoefficientEstimator,
+)
+from datasets import load_dataset
+from tqdm import tqdm
+
+# Initialize estimator
+estimator = TeaCacheCoefficientEstimator(
+    model_path="/path/to/your/model",
+    model_type="YourModel",
+)
+
+# Load diverse prompts (paper recommends ~70 prompts)
+dataset = load_dataset("nateraw/parti-prompts", split="train")
+prompts = dataset["Prompt"][:70]
+
+# Collect data
+for prompt in tqdm(prompts, desc="Collecting data"):
+    estimator.collect_from_prompt(prompt=prompt, num_inference_steps=50)
+
+# Estimate coefficients
+coeffs = estimator.estimate(poly_order=4)
+print(f"Estimated coefficients: {coeffs.tolist()}")
+```
+
+**Data Statistics Guide:**
+
+| Metric | Good Range | Warning Signs |
+|--------|------------|---------------|
+| **Count** | 2000-5000+ | < 1000: too few prompts |
+| **Input Diffs (x)** | 0.01-0.10 | Very small (<0.001): model may not modulate properly |
+| **Output Diffs (y)** | Should correlate with x | No correlation: check extractor |
+| **Coefficient magnitude** | -1e6 to 1e6 | > 1e8: numerical instability |
+
+---
+
+## Testing
+
+After adding TeaCache support, test with:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="your-model-name",
+    cache_backend="tea_cache",
+    cache_config={
+        "rel_l1_thresh": 0.2,
+        "coefficients": [1.33e6, -1.69e5, 7.95e3, -1.64e2, 1.26],  # Your coefficients
+    }
+)
+
+images = omni.generate(
+    "a beautiful landscape",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+**Verify:**
+
+1. **Check logs** - Look for TeaCache initialization messages
+2. **Compare performance** - Measure speedup vs baseline (expect 1.5x-2.0x)
+3. **Verify output quality** - Visually compare cached vs uncached outputs (should be nearly identical)
+
+See more detailed examples in [user guide for teacache](../../user_guide/diffusion/teacache.md).
+
+---
+
+## Troubleshooting
+
+### Issue: "Unknown model type"
+
+**Symptoms:** Error message indicating the model type is not recognized when enabling TeaCache.
+
+**Causes & Solutions:**
+
+- **Extractor not registered:**
+
+**Problem:** The transformer class name doesn't exist in `EXTRACTOR_REGISTRY`.
+
+**Solution:** Check the class name and add to registry:
+```python
+# Check transformer class name
+print(pipeline.transformer.__class__.__name__)
+
+# Add to EXTRACTOR_REGISTRY
+EXTRACTOR_REGISTRY["YourTransformer2DModel"] = extract_your_context
+```
+
+- **Transformer class name mismatch:**
+
+**Solution:** Ensure the registry key matches exactly with `module.__class__.__name__`.
+
+### Issue: "Cannot find coefficients"
+
+**Symptoms:** Error when initializing TeaCache about missing model coefficients.
+
+**Causes & Solutions:**
+
+- **Missing coefficients in config:**
+
+**Solution:** Add coefficients to `_MODEL_COEFFICIENTS` in `config.py`, or pass custom coefficients:
+```python
+omni = Omni(
+    model="your-model",
+    cache_backend="tea_cache",
+    cache_config={"coefficients": [1.0, -0.5, 0.1, -0.01, 0.001]}
+)
+```
+
+### Issue: Quality Degradation
+
+**Symptoms:** Output images look noticeably different or have artifacts compared to baseline.
+
+**Causes & Solutions:**
+
+- **Threshold too high:**
+
+**Problem:** `rel_l1_thresh` is too aggressive, causing cache reuse when outputs differ significantly.
+
+**Solution:** Lower the threshold:
+```python
+cache_config={"rel_l1_thresh": 0.1}  # Try 0.1-0.2
+```
+
+- **Coefficients not tuned:**
+
+**Solution:** Estimate model-specific coefficients using the coefficient estimation process described above.
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Model | Path | Pattern | Notes |
+|-------|------|---------|-------|
+| **Qwen-Image** | `vllm_omni/diffusion/cache/teacache/extractors.py` | Dual-stream | `extract_qwen_context` |
+| **Bagel** | `vllm_omni/diffusion/cache/teacache/extractors.py` | Omni model | `extract_bagel_context` |
+| **TeaCache Core** | `vllm_omni/diffusion/cache/teacache/` | Base implementation | Hook and config |
+| **Coefficient Estimator** | `vllm_omni/diffusion/cache/teacache/coefficient_estimator.py` | Estimation tool | Adapter pattern |
+
+---
+
+## Summary
+
+Adding TeaCache support:
+
+1. ✅ **Write extractor** - Create function returning `CacheContext` with model-specific preprocessing
+2. ✅ **Register extractor** - Add to `EXTRACTOR_REGISTRY` with transformer class name
+3. ✅ **Add coefficients** - Add polynomial coefficients to `_MODEL_COEFFICIENTS`
+4. ✅ **Test** - Verify with `cache_backend="tea_cache"`

docs/design/feature/tensor_parallel.md

@@ -0,0 +1,279 @@
+# Tensor Parallel
+
+This section describes how to add Tensor Parallel (TP) to a diffusion transformer model. We use the Z-Image transformer as the reference implementation.
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Step-by-Step Implementation](#step-by-step-implementation)
+- [Testing](#testing)
+- [Troubleshooting](#troubleshooting)
+- [Reference Implementations](#reference-implementations)
+- [Summary](#summary)
+
+---
+
+## Overview
+
+### What is Tensor Parallel?
+
+Tensor Parallel (TP) is a model parallelism technique that **shards model weights** across multiple GPUs. Each GPU holds only a portion of the model's parameters and computes only part of each layer's output.
+
+Diffusion transformers contain large attention and MLP layers. We can use Tensor Parallel to shard the model dimension across multiple GPUs, allowing larger models to fit in memory while achieving near-linear speedup.
+
+### Architecture
+
+The Tensor Parallel implementation relies vLLM's Parallel Layers:
+
+[vLLM Parallel Layers API Reference](https://docs.vllm.ai/en/latest/contributing/model/basic/?h=column#3-optional-implement-tensor-parallelism-and-quantization-support)
+
+**Parallel Layer Types:**
+
+| Layer Type | Purpose | Weight Partitioning |
+|------------|---------|---------------------|
+| `ColumnParallelLinear` | First FFN layer, separated QKV | Columns (output dimension) |
+| `RowParallelLinear` | Second FFN layer, attention output | Rows (input dimension) |
+| `QKVParallelLinear` | Multi-head/grouped-query attention QKV | Handles head replication automatically |
+| `ReplicatedLinear` | Layers that shouldn't be sharded | No partitioning (replicated) |
+
+---
+
+## Step-by-Step Implementation
+
+
+### Step 1: Identify Linear Layers
+
+Find all `nn.Linear` layers in your transformer that need to be sharded.
+
+**Key questions:**
+- Which layers should be column parallel (weight split by columns)?
+- Which layers should be row parallel (weight split by rows)?
+
+### Step 2: Replace Linear Layers with Parallel Equivalents
+
+Replace `nn.Linear` with parallel layers from `vllm.model_executor.layers.linear`.
+
+**Example (MLP Block - Up-Down Pattern):**
+
+```python
+class FeedForward(nn.Module):
+    def __init__(self, dim: int, hidden_dim: int):
+        super().__init__()
+        # Column parallel: weight split by columns [hidden_dim/N, dim]
+        self.w1 = ColumnParallelLinear(
+            dim,
+            hidden_dim,
+            bias=False,
+            return_bias=False,
+        )
+        self.act = nn.GELU()
+
+        self.w2 = RowParallelLinear(
+            hidden_dim,
+            dim,
+            bias=False,
+            input_is_parallel=True,  # Input already sharded from w1
+            return_bias=False,
+        )
+
+    def forward(self, x):
+        # x: [batch, seq, dim] (replicated on all GPUs)
+        # w1 outputs sharded [batch, seq, hidden_dim/N]
+        x = self.w1(x)
+        # act operates on sharded tensors (no communication)
+        x = self.act(x)
+        # w2 outputs full dim [batch, seq, dim] via all-reduce
+        x = self.w2(x)
+        return x
+```
+
+**Example (Attention - QKV-Out Pattern):**
+
+```python
+from vllm_omni.diffusion.attention.layer import Attention
+class YourModelAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int, num_kv_heads: int):
+        super().__init__()
+        self.head_dim = dim // num_heads
+
+        # Column parallel: QKV weight split by columns
+        # Each GPU gets num_heads/N heads
+        self.to_qkv = QKVParallelLinear(
+            hidden_size=dim,
+            head_size=self.head_dim,
+            total_num_heads=num_heads,
+            total_num_kv_heads=num_kv_heads,
+            bias=False,
+            return_bias=False,
+        )
+
+        # Row parallel: output weight split by rows
+        self.to_out = RowParallelLinear(
+            dim,
+            dim,
+            bias=False,
+            input_is_parallel=True,  # Input sharded from attention
+            return_bias=False,
+        )
+
+        self.attn = Attention(
+            num_heads=self.to_qkv.num_heads, # Each GPU gets num_heads/N heads
+            head_size=self.head_dim,
+            softmax_scale=1.0 / (self.head_dim**0.5),
+            causal=False,
+            num_kv_heads=self.to_qkv.num_kv_heads,
+        )
+
+    def forward(self, x):
+        # x: [batch, seq, dim] (replicated)
+        # to_qkv outputs sharded [batch, seq, (q+k+v) * head_dim/N]
+        qkv = self.to_qkv(x)
+        # Split into Q, K, V (each sharded on heads)
+        q, k, v = qkv.split([...], dim=-1)
+        # Attention computed independently on each GPU
+        out = self.attn(q, k, v)
+        # to_out all-reduces to full dim
+        out = self.to_out(out)
+        return out
+```
+
+**Key Points:**
+
+- `ColumnParallelLinear` → `RowParallelLinear` is the standard pairing
+- Set `input_is_parallel=True` on `RowParallelLinear` when input comes from `ColumnParallelLinear`
+- Use `QKVParallelLinear` for attention projections (handles head replication automatically)
+
+### Step 3: Validate TP Constraints
+
+For correct TP operation, these dimensions **must be divisible** by `tensor_parallel_size`:
+
+| Dimension | Reason | Example Error |
+|-----------|--------|---------------|
+| `num_heads` | Heads sharded by QKVParallelLinear | `num_heads=30, tp=4` ❌ (30 % 4 ≠ 0) |
+| `num_kv_heads` | KV heads sharded by QKVParallelLinear | `num_kv_heads=30, tp=4` ❌ (30 % 4 ≠ 0) |
+
+---
+
+## Testing
+
+After adding Tensor Parallel support, test with:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+parallel_config = DiffusionParallelConfig(tensor_parallel_size=2)
+omni = Omni(model="your-model-name", parallel_config=parallel_config)
+
+output = omni.generate(
+    "a cup of coffee on the table",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+**Or via command line:**
+
+```bash
+cd examples/offline_inference/text_to_image
+python text_to_image.py \
+    --model Your-org/your-model \
+    --prompt "a cup of coffee on the table" \
+    --negative-prompt "ugly, unclear" \
+    --cfg-scale 4.0 \
+    --num-inference-steps 50 \
+    --output "tp_enabled.png" \
+    --tensor-parallel-size 2
+```
+
+**Verify:**
+
+1. Check the `e2e_time_ms` in the log for speedup
+2. Compare generated image quality with TP disabled
+3. Verify memory usage is reduced proportionally
+4. Record comparison results in your PR
+
+---
+
+## Troubleshooting
+
+### Issue: TP not activating
+
+**Symptoms:** Model runs on single GPU, no memory savings or speedup.
+
+**Causes & Solutions:**
+
+- **Still using `nn.Linear`:**
+
+**Problem:** Linear layers not replaced with parallel equivalents.
+
+**Solution:** Replace with parallel layers:
+```python
+# ❌ BAD
+self.proj = nn.Linear(dim, dim)
+
+# ✅ GOOD
+self.proj = RowParallelLinear(dim, dim, input_is_parallel=True)
+```
+
+### Issue: Dimension mismatch errors
+
+**Symptoms:** `RuntimeError: shape mismatch` during forward pass.
+
+**Causes & Solutions:**
+
+- **Missing `input_is_parallel=True`:**
+
+**Problem:** RowParallelLinear expects sharded input but receives full tensor.
+
+**Solution:** Set `input_is_parallel=True` when input comes from ColumnParallelLinear:
+```python
+# ✅ GOOD: Correct pairing
+self.w1 = ColumnParallelLinear(dim, hidden_dim, return_bias=False,)
+self.w2 = RowParallelLinear(
+    hidden_dim,
+    dim,
+    input_is_parallel=True,  # Input sharded from w1
+    return_bias=False,
+)
+```
+
+- **Incorrect split dimensions:**
+
+**Problem:** QKV split sizes don't match sharded dimensions.
+
+**Solution:** Use `self.to_qkv.num_heads` (local heads per GPU):
+```python
+# ❌ BAD: Uses total heads
+q_size = self.total_num_heads * self.head_dim
+
+# ✅ GOOD: Uses local heads
+q_size = self.to_qkv.num_heads * self.head_dim
+```
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Model | Path | Pattern | Notes |
+|-------|------|---------|-------|
+| **Z-Image** | `vllm_omni/diffusion/models/z_image/z_image_transformer.py` | Standard TP | Full implementation with validation |
+| **FLUX** | `vllm_omni/diffusion/models/flux/flux_transformer.py` | Dual-stream | Image + text streams |
+| **Qwen-Image** | `vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py` | Standard TP | With RoPE |
+| **TP Tests** | `tests/e2e/offline_inference/test_zimage_parallelism.py` | E2E testing | TP correctness and performance |
+| **Constraint Tests** | `tests/diffusion/models/z_image/test_zimage_tp_constraints.py` | Unit testing | Validation logic |
+
+---
+
+## Summary
+
+Adding Tensor Parallel support to a transformer:
+
+1. ✅ **Identify linear layers** - Which layers should be sharded?
+2. ✅ **Replace with parallel layers** - Use QKVParallelLinear, ColumnParallelLinear, RowParallelLinear
+3. ✅ **Validate TP constraints** - Ensure dimensions divisible by TP size
+4. ✅ **Test** - Verify with `tensor_parallel_size=N`, check memory, speed, and quality

docs/features/custom_pipeline.md

@@ -0,0 +1,151 @@
+# Custom Pipeline Extension Guide
+
+Transformer already support Custom Pipeline via
+https://github.com/huggingface/diffusers/blob/main/docs/source/en/using-diffusers/custom_pipeline_overview.md
+
+This guide demonstrates how to use the newly added features for extending vLLM-Omni's diffusion pipeline with custom functionality.
+
+## Overview
+
+Three main features enable custom pipeline extension:
+
+1. **`WorkerWrapperBase`**: A wrapper class that enables dynamic worker extension with custom functionality
+2. **`load_format`**: A parameter that controls how diffusion models are loaded, including support for custom pipelines
+3. **`CustomPipelineWorkerExtension`**: An extension class that enables pipeline re-initialization with custom implementations
+
+## Features
+
+### WorkerWrapperBase
+
+`WorkerWrapperBase` is a wrapper class that creates `DiffusionWorker` instances with optional extension support. It enables dynamic inheritance, allowing you to add custom methods and functionality to workers without modifying the base worker class.
+
+**Key capabilities:**
+- Dynamic worker class extension via `worker_extension_cls`
+- Support for custom pipeline initialization via `custom_pipeline_args`
+- Method delegation to underlying worker
+- Attribute access forwarding
+
+**Location:** `vllm_omni/diffusion/worker/diffusion_worker.py`
+
+### load_format Parameter
+
+The `load_format` parameter controls how diffusion models are loaded. It supports the following values:
+
+- **`"default"`**: Standard model loading using the model registry (default behavior)
+- **`"custom_pipeline"`**: Load a custom pipeline class specified by `custom_pipeline_name`
+- **`"dummy"`**: Skip model loading (useful for testing or when pipeline will be initialized separately)
+
+**Location:** `vllm_omni/diffusion/model_loader/diffusers_loader.py`
+
+### CustomPipelineWorkerExtension
+
+`CustomPipelineWorkerExtension` is a mixin class that extends `DiffusionWorker` with the ability to re-initialize the pipeline with a custom implementation.
+
+**Key method:**
+- `re_init_pipeline(custom_pipeline_args)`: Re-initializes the pipeline with custom arguments, properly cleaning up the old pipeline first
+
+**Location:** `vllm_omni/diffusion/worker/diffusion_worker.py`
+
+## Usage Example
+
+### Step 1: Create a Custom Pipeline
+
+Create a custom pipeline class that extends an existing pipeline. In this example, we extend `QwenImageEditPipeline` to add trajectory tracking:
+
+```python
+# custom_pipeline.py
+from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
+from vllm_omni.diffusion.models.qwen_image.pipeline_qwen_image_edit import QwenImageEditPipeline
+import torch
+
+class CustomPipeline(QwenImageEditPipeline):
+    def __init__(self, *, od_config: OmniDiffusionConfig, prefix: str = ""):
+        super().__init__(od_config=od_config, prefix=prefix)
+
+    def forward(self, req, prompt=None, negative_prompt=None, **kwargs):
+        # Call parent's forward to get normal output
+        output = super().forward(req=req, prompt=prompt, negative_prompt=negative_prompt, **kwargs)
+
+        # Add custom trajectory data
+        actual_num_steps = req.sampling_params.num_inference_steps or kwargs.get('num_inference_steps', 50)
+        output.trajectory_timesteps = torch.linspace(1000, 0, actual_num_steps, dtype=torch.float32)
+        output.trajectory_latents = torch.randn(actual_num_steps, 1, 16, 64, 64, dtype=torch.float32)
+
+        return output
+```
+
+### Step 2: Use the Custom Pipeline with Omni
+
+Initialize the `Omni` engine with custom pipeline configuration:
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+# Initialize with custom pipeline
+omni = Omni(
+    model="Qwen/Qwen-Image-Edit",
+    diffusion_load_format="dummy",  # Skip initial loading
+    custom_pipeline_args={
+        "pipeline_class": "custom_pipeline.CustomPipeline"
+    },
+)
+
+# Generate with the custom pipeline
+outputs = omni.generate(
+    ...
+)
+
+# Access custom trajectory data
+output = outputs[0].request_output[0]
+print(f"Trajectory timesteps shape: {output.metrics['trajectory_timesteps'].shape}")
+print(f"Trajectory latents shape: {output.latents.shape}")
+```
+
+### Step 3: Run the Example
+
+The example provided in this directory demonstrates the complete workflow:
+
+```bash
+cd examples/offline_inference/custom_pipeline/image_to_image
+
+# Run with custom pipeline
+python image_edit.py \
+    --model Qwen/Qwen-Image-Edit-2511 \
+    --image cherry_blossom.jpg \
+    --prompt "Let this mascot dance under the moon, surrounded by floating stars" \
+    --output output_image_edit.png \
+    --num_inference_steps 10
+```
+
+## Advanced Usage
+
+### Custom Worker Extension
+
+You can create custom worker extensions to add new methods beyond pipeline re-initialization:
+
+```python
+from typing import Any
+from vllm_omni.diffusion.worker.diffusion_worker import CustomPipelineWorkerExtension
+
+class MyCustomExtension(CustomPipelineWorkerExtension):
+    def custom_method(self):
+        """Your custom worker method."""
+        return "custom_result"
+
+    def another_method(self, data: Any):
+        """Another custom method."""
+        # Access worker internals via self
+        return self.model_runner.some_operation(data)
+
+omni = Omni(
+    model="Qwen/Qwen-Image-Edit",
+    diffusion_load_format="dummy",
+    custom_pipeline_args={
+        "pipeline_class": "custom_pipeline.CustomPipeline"
+    },
+    worker_extension_cls=MyCustomExtension,
+    # Note: worker_extension_cls is an internal parameter
+    # CustomPipelineWorkerExtension will automatically init pipeline when custom_pipeline_args is provided
+)
+```

docs/getting_started/installation/README.md

@@ -5,4 +5,5 @@ vLLM-Omni supports the following hardware platforms:
 - [GPU](gpu.md)
     - [NVIDIA CUDA](gpu.md)
     - [AMD ROCm](gpu.md)
+    - [Intel XPU](gpu.md)
 - [NPU](npu.md)

docs/getting_started/installation/gpu.md

@@ -18,6 +18,10 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr
 
     --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:requirements"
 
+=== "Intel XPU"
+
+    --8<-- "docs/getting_started/installation/gpu/xpu.inc.md:requirements"
+
 ## Set up using Python
 
 ### Create a new Python environment
@@ -32,11 +36,14 @@ Note: Pre-built wheels are currently only available for vLLM-Omni 0.11.0rc1, 0.1
 
     --8<-- "docs/getting_started/installation/gpu/cuda.inc.md:pre-built-wheels"
 
-
 === "AMD ROCm"
 
     --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:pre-built-wheels"
 
+=== "Intel XPU"
+
+    --8<-- "docs/getting_started/installation/gpu/xpu.inc.md:pre-built-wheels"
+
 [](){ #build-from-source }
 
 ### Build wheel from source
@@ -49,6 +56,10 @@ Note: Pre-built wheels are currently only available for vLLM-Omni 0.11.0rc1, 0.1
 
     --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:build-wheel-from-source"
 
+=== "Intel XPU"
+
+    --8<-- "docs/getting_started/installation/gpu/xpu.inc.md:build-wheel-from-source"
+
 ## Set up using Docker
 
 ### Pre-built images
@@ -61,8 +72,16 @@ Note: Pre-built wheels are currently only available for vLLM-Omni 0.11.0rc1, 0.1
 
     --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:pre-built-images"
 
+=== "Intel XPU"
+
+    --8<-- "docs/getting_started/installation/gpu/xpu.inc.md:pre-built-images"
+
 ### Build your own docker image
 
 === "AMD ROCm"
 
     --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:build-docker"
+
+=== "Intel XPU"
+
+    --8<-- "docs/getting_started/installation/gpu/xpu.inc.md:build-docker"

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==0.15.0 --torch-backend=auto
+uv pip install vllm --torch-backend=auto
 ```
 
 #### Installation of vLLM-Omni
@@ -34,10 +34,10 @@ uv pip install vllm-omni
 # --8<-- [start:build-wheel-from-source]
 
 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.15.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.16.0 release version of the library
 
 ```bash
-uv pip install vllm==0.15.0 --torch-backend=auto
+uv pip install vllm==0.16.0 --torch-backend=auto
 ```
 
 The release 0.14.0 of vLLM is based on PyTorch 2.9.0 which requires CUDA 12.9 environment.
@@ -56,11 +56,11 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.15.0
+git checkout v0.16.0
 ```
 Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
 ```bash
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0-cp38-abi3-manylinux_2_31_x86_64.whl
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.16.0/vllm-0.16.0-cp38-abi3-manylinux_2_31_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash

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==0.14.0+rocm700 --extra-index-url https://wheels.vllm.ai/rocm/0.14.0/rocm700
+uv pip install vllm==0.16.0+rocm700 --extra-index-url https://wheels.vllm.ai/rocm/0.16.0/rocm700
 ```
 
 #### Installation of vLLM-Omni
@@ -34,13 +34,13 @@ uv pip install onnxruntime-rocm sox
 # --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 0.14.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.16.0 release version of the library
 
 ```bash
-uv pip install vllm==0.14.0+rocm700 --extra-index-url https://wheels.vllm.ai/rocm/0.14.0/rocm700
+uv pip install vllm==0.16.0+rocm700 --extra-index-url https://wheels.vllm.ai/rocm/0.16.0/rocm700
 ```
 
-The release 0.14.0 of vLLM requires ROCm 7.0 environment.
+The release 0.16.0 of vLLM requires ROCm 7.0 environment.
 
 #### 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 v0.14.0
+git checkout v0.16.0
 python3 -m pip install -r requirements/rocm.txt
 python3 setup.py develop
 ```

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

@@ -0,0 +1,48 @@
+# --8<-- [start:requirements]
+
+- GPU: Validated on Intel® Arc™ B-Series.
+
+# --8<-- [end:requirements]
+# --8<-- [start:set-up-using-python]
+
+vLLM-Omni currently recommends using the Docker image setup steps below.
+
+# --8<-- [start:pre-built-wheels]
+
+# --8<-- [end:pre-built-wheels]
+
+# --8<-- [start:build-wheel-from-source]
+
+# --8<-- [end:build-wheel-from-source]
+
+# --8<-- [start:build-docker]
+
+#### Build docker image
+
+```bash
+DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.xpu -t vllm-omni-xpu --shm-size=4g .
+```
+
+#### Launch the docker image
+
+##### Launch with OpenAI API Server
+
+```
+docker run -it -d --shm-size 10g \
+  --name {container_name} \
+  --net=host \
+  --ipc=host \
+  --privileged \
+  -v /dev/dri/by-path:/dev/dri/by-path \
+  --device /dev/dri:/dev/dri \
+  -v ~/.cache/huggingface:/root/.cache/huggingface \
+  --env "HF_TOKEN=$HF_TOKEN" \
+  vllm-omni-xpu \
+  --model Qwen/Qwen2.5-Omni-3B --port 8091
+```
+
+# --8<-- [end:build-docker]
+
+# --8<-- [start:pre-built-images]
+
+# --8<-- [end:pre-built-images]

docs/getting_started/installation/npu.md

@@ -24,6 +24,10 @@ vLLM-Omni supports NPU through the vLLM Ascend Plugin (vllm-ascend). This is a c
 
 ### Build wheel from source
 
-=== "NPU"
+=== "NPU release"
 
-    --8<-- "docs/getting_started/installation/npu/npu.inc.md:installation"
+    --8<-- "docs/getting_started/installation/npu/npu.inc.md:installation-release"
+
+=== "NPU from main"
+
+    --8<-- "docs/getting_started/installation/npu/npu.inc.md:installation-main"

docs/getting_started/installation/npu/npu.inc.md

@@ -3,14 +3,11 @@
 For detailed hardware and software requirements, please refer to the [vllm-ascend installation documentation](https://docs.vllm.ai/projects/ascend/en/latest/installation.html).
 
 # --8<-- [end:requirements]
-# --8<-- [start:installation]
+# --8<-- [start:installation-release]
 
 The recommended way to use vLLM-Omni on NPU is through the vllm-ascend pre-built Docker images:
 
 ```bash
-# Update DEVICE according to your NPUs (/dev/davinci[0-7])
-export DEVICE0=/dev/davinci0
-export DEVICE1=/dev/davinci1
 # Update the vllm-ascend image
 # Atlas A2:
 # export IMAGE=quay.io/ascend/vllm-ascend:v0.14.0rc1
@@ -20,8 +17,10 @@ export IMAGE=quay.io/ascend/vllm-ascend:v0.14.0rc1
 docker run --rm \
     --name vllm-omni-npu \
     --shm-size=1g \
-    --device $DEVICE0 \
-    --device $DEVICE1 \
+    --device /dev/davinci0 \
+    --device /dev/davinci1 \
+    --device /dev/davinci2 \
+    --device /dev/davinci3 \
     --device /dev/davinci_manager \
     --device /dev/devmm_svm \
     --device /dev/hisi_hdc \
@@ -34,13 +33,18 @@ docker run --rm \
     -p 8000:8000 \
     -it $IMAGE bash
 
+# Because vllm-ascend will release v0.16.0rc1 after vllm-omni 0.16.0,
+# we have to pin vllm-ascend at the current commit.
+cd /vllm-workspace/vllm-ascend
+git checkout e2175d9c7e62b437391dfee996b1375674ba7c18
+pip install -v -e .
+
 # Inside the container, install vLLM-Omni from source
 cd /vllm-workspace
-git clone -b v0.14.0 https://github.com/vllm-project/vllm-omni.git
+git clone -b v0.16.0 https://github.com/vllm-project/vllm-omni.git
 
 cd vllm-omni
-VLLM_OMNI_TARGET_DEVICE=npu pip install -v -e .
-# OR pip install -v -e . --no-build-isolation
+pip install -v -e .
 export VLLM_WORKER_MULTIPROC_METHOD=spawn
 ```
 
@@ -48,13 +52,42 @@ The default workdir is `/workspace`, with vLLM, vLLM-Ascend and vLLM-Omni code p
 
 For other installation methods (pip installation, building from source, custom Docker builds), please refer to the [vllm-ascend installation guide](https://docs.vllm.ai/projects/ascend/en/latest/installation.html).
 
-We are keeping [issue #997](https://github.com/vllm-project/vllm-omni/issues/997) up to date with the aligned versions of vLLM, vLLM-Ascend, and vLLM-Omni, and also outlining the Q1 roadmap there.
+We are keeping [issue #886](https://github.com/vllm-project/vllm-omni/issues/886) up to date with the aligned versions of vLLM, vLLM-Ascend, and vLLM-Omni, and also outlining the Q1 roadmap there.
 
-# --8<-- [end:installation]
+# --8<-- [end:installation-release]
+
+# --8<-- [start:installation-main]
+
+You can also build vLLM-Omni from the latest main branch if you want to use the latest features or bug fixes. (But sometimes it will break for a while. You can check [issue #886](https://github.com/vllm-project/vllm-omni/issues/886) for the status of the latest commit of vLLM-Omni main branch on NPU.)
+
+```bash
+# Pin vLLM version to 0.16.0
+cd /vllm-workspace/vllm
+git pull origin main
+git fetch origin --tags
+git checkout v0.16.0
+VLLM_TARGET_DEVICE=empty pip install -v -e .
+
+# Because vllm-ascend has not yet entered continuous development and has not been officially released, we need to pin it to a specific commit. Please note that this commit may change over time.
+cd ../vllm-ascend
+git pull origin main
+git fetch origin --tags
+git checkout e2175d9c7e62b437391dfee996b1375674ba7c18
+pip install -v -e .
+
+# Install vLLM-Omni from the latest main branch
+cd ../vllm-omni
+git clone https://github.com/vllm-project/vllm-omni.git
+pip install -v -e . --no-build-isolation
+# or VLLM_OMNI_TARGET_DEVICE=npu pip install -v -e .
+export VLLM_WORKER_MULTIPROC_METHOD=spawn
+```
+
+# --8<-- [end:installation-main]
 
 # --8<-- [start:pre-built-images]
 
-`vllm-ascend` offers Docker images for deployment. You can just pull the **prebuilt image** from the image repository [ascend/vllm-ascend](https://quay.io/repository/ascend/vllm-ascend?tab=tags) and run it with bash.
+vLLM-Omni offers Docker images for Ascend NPU deployment. You can just pull the **prebuilt image** from the image repository [ascend/vllm-omni](https://quay.io/repository/ascend/vllm-omni?tab=tags) and run it with bash.
 
 Supported images as following.
 
@@ -63,16 +96,14 @@ Supported images as following.
 | image-tag | Atlas A2 | Ubuntu |
 | image-tag-a3 | Atlas A3 | Ubuntu |
 
-vLLM-Omni offers Docker images for Ascend NPU deployment. You can just pull the **prebuilt image** from the image repository [ascend/vllm-ascend](https://quay.io/repository/ascend/vllm-ascend?tab=tags) and run it with bash.
-
-Here's an example deployment command that has been verified on 6 x NPUs:
+Here's an example deployment command that has been verified on 4 x NPUs:
 
 ```bash
 # Atlas A2:
-# export IMAGE=quay.io/ascend/vllm-omni:v0.14.0
+# export IMAGE=quay.io/ascend/vllm-omni:v0.16.0
 # Atlas A3:
-# export IMAGE=quay.io/ascend/vllm-omni:v0.14.0-a3
-export IMAGE=quay.io/ascend/vllm-omni:v0.14.0
+# export IMAGE=quay.io/ascend/vllm-omni:v0.16.0-a3
+export IMAGE=quay.io/ascend/vllm-omni:v0.16.0
 docker run --rm \
     --name vllm-omni-npu \
     --shm-size=1g \
@@ -80,8 +111,6 @@ docker run --rm \
     --device /dev/davinci1 \
     --device /dev/davinci2 \
     --device /dev/davinci3 \
-    --device /dev/davinci4 \
-    --device /dev/davinci5 \
     --device /dev/davinci_manager \
     --device /dev/devmm_svm \
     --device /dev/hisi_hdc \

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==0.15.0 --torch-backend=auto
+uv pip install vllm==0.16.0 --torch-backend=auto
 
 # On ROCm
-uv pip install vllm==0.15.0 --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
+uv pip install vllm==0.16.0 --extra-index-url https://wheels.vllm.ai/rocm/0.16.0/rocm700
 
 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni

docs/mkdocs/hooks/generate_argparse.py

@@ -0,0 +1,258 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+import ast
+import logging
+import re
+import sys
+from argparse import SUPPRESS, Action, ArgumentParser, HelpFormatter, _ArgumentGroup
+from collections.abc import Iterable
+from importlib.machinery import ModuleSpec
+from pathlib import Path
+from typing import Literal
+from unittest.mock import MagicMock
+
+from pydantic_core import core_schema
+
+
+class _FlexibleArgumentParser(ArgumentParser):
+    """Fallback parser for docs when vllm is unavailable.
+
+    Accepts the 'deprecated' kwarg used by vllm CLI and emits warnings
+    if a deprecated argument is actually provided.
+    """
+
+    _deprecated: set[Action] = set()
+    _deprecated_warned: set[str] = set()
+
+    if sys.version_info < (3, 13):
+
+        def parse_known_args(self, args=None, namespace=None):
+            namespace, args = super().parse_known_args(args, namespace)
+            for action in _FlexibleArgumentParser._deprecated:
+                if (
+                    hasattr(namespace, dest := action.dest)
+                    and getattr(namespace, dest) != action.default
+                    and dest not in _FlexibleArgumentParser._deprecated_warned
+                ):
+                    _FlexibleArgumentParser._deprecated_warned.add(dest)
+                    logger.warning("argument '%s' is deprecated", dest)
+            return namespace, args
+
+    def add_argument(self, *args, **kwargs):
+        deprecated = kwargs.pop("deprecated", False)
+        action = super().add_argument(*args, **kwargs)
+        if deprecated:
+            _FlexibleArgumentParser._deprecated.add(action)
+        return action
+
+    class _FlexibleArgumentGroup(_ArgumentGroup):
+        def add_argument(self, *args, **kwargs):
+            deprecated = kwargs.pop("deprecated", False)
+            action = super().add_argument(*args, **kwargs)
+            if deprecated:
+                _FlexibleArgumentParser._deprecated.add(action)
+            return action
+
+    def add_argument_group(self, *args, **kwargs):
+        group = self._FlexibleArgumentGroup(self, *args, **kwargs)
+        self._action_groups.append(group)
+        return group
+
+
+logger = logging.getLogger("mkdocs")
+
+# Define root and doc output directories
+ROOT_DIR = Path(__file__).parent.parent.parent.parent
+ARGPARSE_DOC_DIR = ROOT_DIR / "docs/generated/argparse_omni"
+
+# Ensure the repo root is in sys.path for dynamic imports
+sys.path.insert(0, str(ROOT_DIR))
+
+
+class PydanticMagicMock(MagicMock):
+    """`MagicMock` that's able to generate pydantic-core schemas and avoids _mock_name=None errors."""
+
+    def __init__(self, *args, **kwargs):
+        name = kwargs.pop("name", None)
+        super().__init__(*args, **kwargs)
+        self.__spec__ = ModuleSpec(name, None)
+        self._mock_name = name or "mock"
+
+    def __get_pydantic_core_schema__(self, source_type, handler):
+        return core_schema.any_schema()
+
+
+# --- Static extraction for CLI argument docs ---
+
+
+def extract_omni_serve_subparser_init():
+    """
+    Statically parse vllm_omni/entrypoints/cli/serve.py to extract the subparser_init method
+    and return a callable that adds arguments to a parser. This avoids import and mock issues.
+    """
+    serve_path = ROOT_DIR / "vllm_omni" / "entrypoints" / "cli" / "serve.py"
+    with open(serve_path, encoding="utf-8") as f:
+        source = f.read()
+    tree = ast.parse(source, filename=str(serve_path))
+    # Find class OmniServeCommand
+    for node in tree.body:
+        if isinstance(node, ast.ClassDef) and node.name == "OmniServeCommand":
+            for item in node.body:
+                if isinstance(item, ast.FunctionDef) and item.name == "subparser_init":
+                    # We'll exec this function body in a dummy context
+                    func_code = ast.Module(body=[item], type_ignores=[])
+                    code = compile(func_code, filename=str(serve_path), mode="exec")
+                    # Prepare dummy context
+                    local_vars = {}
+
+                    # Provide a dummy subparsers with add_parser returning our parser
+                    class DummySubparsers:
+                        def add_parser(self, name, **kwargs):
+                            return _FlexibleArgumentParser(prog=name)
+
+                    dummy_subparsers = DummySubparsers()
+                    # Provide globals for exec
+                    # Extract DESCRIPTION from the source file (assume it's a triple-quoted string at the top)
+                    m = re.search(r'DESCRIPTION\s*=\s*([ru]?""".*?""")', source, re.DOTALL)
+                    DESCRIPTION = m.group(1)[3:-3] if m else ""
+                    exec_globals = {
+                        "_FlexibleArgumentParser": _FlexibleArgumentParser,
+                        "FlexibleArgumentParser": _FlexibleArgumentParser,
+                        "make_arg_parser": lambda parser: parser,  # no-op for doc
+                        "VLLM_SUBCMD_PARSER_EPILOG": "",
+                        "logger": logger,
+                        "DummySubparsers": DummySubparsers,
+                        "argparse": __import__("argparse"),
+                        "json": __import__("json"),
+                        "DESCRIPTION": DESCRIPTION,
+                    }
+                    exec(code, exec_globals, local_vars)
+                    # Get the function
+                    subparser_init = local_vars["subparser_init"]
+
+                    # Return a callable that mimics OmniServeCommand().subparser_init
+                    def parser_factory():
+                        class DummySelf:
+                            name = "serve"
+
+                        return subparser_init(DummySelf(), dummy_subparsers)
+
+                    return parser_factory
+    raise RuntimeError("Could not statically extract OmniServeCommand.subparser_init")
+
+
+OmniServeCommand = type("OmniServeCommand", (), {"subparser_init": staticmethod(extract_omni_serve_subparser_init())})
+
+
+class MarkdownFormatter(HelpFormatter):
+    """Custom formatter that generates markdown for argument groups."""
+
+    def __init__(self, prog: str, starting_heading_level: int = 3):
+        super().__init__(prog, max_help_position=sys.maxsize, width=sys.maxsize)
+
+        self._section_heading_prefix = "#" * starting_heading_level
+        self._argument_heading_prefix = "#" * (starting_heading_level + 1)
+        self._markdown_output = []
+
+    def start_section(self, heading: str):
+        if heading not in {"positional arguments", "options"}:
+            heading_md = f"\n{self._section_heading_prefix} {heading}\n\n"
+            self._markdown_output.append(heading_md)
+
+    def end_section(self):
+        pass
+
+    def add_text(self, text: str):
+        if text:
+            self._markdown_output.append(f"{text.strip()}\n\n")
+
+    def add_usage(self, usage, actions, groups, prefix=None):
+        pass
+
+    def add_arguments(self, actions: Iterable[Action]):
+        for action in actions:
+            if len(action.option_strings) == 0 or "--help" in action.option_strings:
+                continue
+
+            option_strings = f"`{'`, `'.join(action.option_strings)}`"
+            heading_md = f"{self._argument_heading_prefix} {option_strings}\n\n"
+            self._markdown_output.append(heading_md)
+
+            if choices := action.choices:
+                choices = f"`{'`, `'.join(str(c) for c in choices)}`"
+                self._markdown_output.append(f"Possible choices: {choices}\n\n")
+            elif (metavar := action.metavar) and isinstance(metavar, (list, tuple)):
+                metavar = f"`{'`, `'.join(str(m) for m in metavar)}`"
+                self._markdown_output.append(f"Possible choices: {metavar}\n\n")
+
+            if action.help:
+                self._markdown_output.append(f"{action.help}\n\n")
+
+            if (default := action.default) != SUPPRESS:
+                # Make empty string defaults visible
+                if default == "":
+                    default = '""'
+                self._markdown_output.append(f"Default: `{default}`\n\n")
+
+    def format_help(self):
+        """Return the formatted help as markdown."""
+        return "".join(self._markdown_output)
+
+
+# Function to create parser using subparser_init style CLI class
+
+
+def create_parser_subparser_init(subcmd_class):
+    """
+    Create an argparse parser using subparser_init style CLI class, with MarkdownFormatter.
+    Compatible with both AST-extracted static parser_factory and real class instance method.
+    """
+
+    class DummySubparsers:
+        def add_parser(self, name, **kwargs):
+            return _FlexibleArgumentParser(prog=name)
+
+    dummy_subparsers = DummySubparsers()
+    # If subparser_init is a zero-arg static function (AST-extracted), call it directly
+    subparser_init = getattr(subcmd_class, "subparser_init", None)
+    if subparser_init is not None:
+        import inspect
+
+        sig = inspect.signature(subparser_init)
+        if len(sig.parameters) == 0:
+            parser = subparser_init()
+        else:
+            parser = subcmd_class().subparser_init(dummy_subparsers)
+    else:
+        parser = subcmd_class().subparser_init(dummy_subparsers)
+    parser.formatter_class = MarkdownFormatter
+    return parser
+
+
+def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
+    """
+    Entry point for doc generation. Builds doc directory and outputs markdown for each CLI command.
+    """
+    logger.info("Generating vllm-omni argparse documentation")
+    logger.debug("Root directory: %s", ROOT_DIR.resolve())
+    logger.debug("Output directory: %s", ARGPARSE_DOC_DIR.resolve())
+    if not ARGPARSE_DOC_DIR.exists():
+        ARGPARSE_DOC_DIR.mkdir(parents=True)
+
+    # Register all CLI parsers; you can easily add more commands here
+    parsers = {
+        "omni_serve": create_parser_subparser_init(OmniServeCommand),
+        # "another_cmd": create_parser_subparser_init(AnotherCommandClass),
+    }
+
+    for stem, parser in parsers.items():
+        doc_path = ARGPARSE_DOC_DIR / f"{stem}.inc.md"
+        with open(doc_path, "w", encoding="utf-8") as f:
+            f.write(super(type(parser), parser).format_help())
+        logger.info("Argparse generated: %s", doc_path.relative_to(ROOT_DIR))
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO)
+    on_startup("build", False)

docs/models/supported_models.md

@@ -20,6 +20,7 @@ th {
 | `Qwen3OmniMoeForConditionalGeneration` | Qwen3-Omni | `Qwen/Qwen3-Omni-30B-A3B-Instruct` |
 | `Qwen2_5OmniForConditionalGeneration` | Qwen2.5-Omni | `Qwen/Qwen2.5-Omni-7B`, `Qwen/Qwen2.5-Omni-3B` |
 | `BagelForConditionalGeneration` | BAGEL (DiT-only) | `ByteDance-Seed/BAGEL-7B-MoT` |
+| `HunyuanImage3ForCausalMM` | HunyuanImage3.0 (DiT-only) | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` |
 | `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` |
 | `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` |
 | `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` |
@@ -35,10 +36,14 @@ th {
 |`StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` |
 |`Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` |
 |`FluxPipeline` | FLUX.1-dev | `black-forest-labs/FLUX.1-dev` |
+|`OmniGen2Pipeline` | OmniGen2 | `OmniGen2/OmniGen2` |
 |`StableAudioPipeline` | Stable-Audio-Open | `stabilityai/stable-audio-open-1.0` |
 |`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` |
 |`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` |
 |`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-0.6B-Base` |
+|`GlmImageForConditionalGeneration` | GLM-Image | `zai-org/GLM-Image` |
+|`NextStep11Pipeline` | NextStep-1.1 | `stepfun-ai/NextStep-1.1` |
+|`MiMoAudioForConditionalGeneration` | MiMo-Audio-7B-Instruct | `XiaomiMiMo/MiMo-Audio-7B-Instruct` |
 
 
 ## List of Supported Models for NPU
@@ -61,7 +66,10 @@ th {
 | `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2511 | `Qwen/Qwen-Image-Edit-2511` |
 |`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
+| `WanPipeline` | Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` |
+| `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` |
 |`LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` |
+|`LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` |
 |`Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` |
 |`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` |

docs/pr_reviewer.md

@@ -0,0 +1,255 @@
+# VLLM-Omni PR Reviewer
+
+## Overview
+
+The VLLM-Omni PR Reviewer is an automated code review bot powered by GLM-4.7 AI model. It helps maintain code quality by providing intelligent feedback on pull requests.
+
+## Features
+
+- **Intelligent Code Analysis**: Leverages GLM-4.7 for understanding code context and providing meaningful feedback
+- **Comprehensive Reviews**: Covers code quality, architecture, security, testing, and documentation
+- **Structured Output**: Provides well-formatted reviews with clear sections and actionable suggestions
+- **Rate Limiting**: Built-in cooldown mechanism to prevent excessive API usage
+- **Retry Logic**: Automatic retries with exponential backoff for transient API failures
+- **Defensive Parsing**: Robust validation of API responses to handle malformed data
+- **Cost Control**: Only repository members/collaborators/owners can trigger reviews
+
+## How to Use
+
+### Triggering a Review
+
+To trigger an automated PR review, mention the bot in a PR comment:
+
+```
+@vllm-omni-reviewer please review
+```
+
+Or include in your PR description:
+
+```
+@vllm-omni-reviewer
+```
+
+The bot will automatically review your changes and post a detailed comment.
+
+## What Gets Reviewed
+
+- **vLLM Architecture Compatibility**: Ensures changes align with vLLM's design patterns
+- **Multi-modal Integration**: Reviews audio, vision, and text processing implementations
+- **Performance Implications**: Analyzes impact on inference latency and throughput
+- **Code Quality**: Checks Python best practices, type hints, and documentation
+- **Security Considerations**: Identifies potential security vulnerabilities
+- **Testing Coverage**: Recommends additional test cases when needed
+
+## Review Output
+
+The bot posts a structured review comment with:
+
+- **Overview**: Brief summary of the PR's purpose
+- **Critical Issues (Must Fix)**: Blocking issues that need to be addressed
+- **Important Issues (Should Fix)**: Significant concerns that should be resolved
+- **Minor Issues & Suggestions**: Small improvements and optional suggestions
+- **Positive Aspects**: Highlights well-implemented features
+- **Performance Considerations**: Analysis of performance impact
+- **Testing Recommendations**: Suggestions for additional tests
+- **Overall Assessment**: Final recommendation (Approve/Request Changes/Needs Major Work)
+
+## Rate Limiting and Cooldown
+
+The bot includes a cooldown mechanism to prevent excessive API usage:
+
+- **Default cooldown**: 5 minutes between reviews per PR
+- **Configurable**: Can be adjusted via `PR_REVIEWER_COOLDOWN_MINUTES` environment variable
+- **Smart detection**: Checks for previous bot comments before starting a review
+
+If you trigger a review within the cooldown period, the bot will log a message and skip the review.
+
+## Architecture
+
+```
+┌─────────────────┐
+│  PR Comment     │
+│  @vllm-omni-    │
+│  reviewer       │
+└────────┬────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  GitHub Actions Workflow        │
+│  (.github/workflows/            │
+│   pr-reviewer.yml)              │
+│                                 │
+│  - Python 3.11                  │
+│  - requests==2.31.0             │
+│  - pyyaml==6.0.1                │
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  PR Reviewer Script             │
+│  (.github/scripts/              │
+│   pr_reviewer.py)               │
+│                                 │
+│  1. Check cooldown              │
+│  2. Fetch PR details & diff     │
+│  3. Build review prompt         │
+│  4. Call GLM-4.7 API            │
+│     (with retry logic)          │
+│  5. Validate response           │
+│  6. Post review comment         │
+└────────┬────────────────────────┘
+         │
+         ▼
+┌─────────────────────────────────┐
+│  GLM-4.7 API                    │
+│  (open.bigmodel.cn)             │
+└─────────────────────────────────┘
+```
+
+1. **GitHub Actions Workflow** (`.github/workflows/pr-reviewer.yml`): Triggers on @mention
+2. **Python Script** (`.github/scripts/pr_reviewer.py`): Fetches PR data and calls GLM-4.7 API
+3. **GLM-4.7 API**: Provides intelligent code analysis
+
+## Testing
+
+### Testing the PR Reviewer Bot
+
+To test the PR reviewer bot before deploying to production:
+
+1. **Create a test PR** - Make a small, safe change (e.g., documentation update)
+2. **Open the PR** - Create a pull request with a descriptive title
+3. **Trigger the review** - Comment `@vllm-omni-reviewer` on the PR
+4. **Monitor results** - Check the Actions tab for workflow execution logs
+
+### Running Unit Tests
+
+The bot includes comprehensive unit tests that can be run locally:
+
+```bash
+# Run all tests
+pytest .github/tests/test_pr_reviewer.py -v
+
+# Run specific test
+pytest .github/tests/test_pr_reviewer.py::TestCheckTrigger -v
+
+# Run with coverage
+pytest .github/tests/test_pr_reviewer.py --cov=.github/scripts/pr_reviewer.py --cov-report=term-missing
+```
+
+### What to Look For
+
+When testing, verify that:
+- [ ] The workflow triggers on the `@vllm-omni-reviewer` comment
+- [ ] The cooldown mechanism works correctly
+- [ ] The GLM API call completes without errors (with retry if needed)
+- [ ] A review comment is posted to the PR
+- [ ] The review content is meaningful and well-structured
+- [ ] The cost is within the expected range (0.50-5 CNY)
+
+### Safe Test Changes
+
+For testing, consider making these types of safe changes:
+- Documentation updates (like adding this Testing section)
+- Comment improvements
+- README enhancements
+- Non-functional file additions
+
+### Example Test PR
+
+A good test PR might:
+- Update a documentation file
+- Add explanatory comments
+- Improve code formatting
+- Fix a minor typo
+
+These changes are safe to merge if the test is successful and won't affect functionality.
+
+## Troubleshooting
+
+### Bot Doesn't Respond
+
+1. **Check permissions** - Verify you have Owner/Member/Collaborator access
+2. **Check Actions tab** - Look for workflow execution and view logs
+3. **Check cooldown** - If another review was posted recently, wait for the cooldown period
+4. **Check API key** - Ensure `GLM_API_KEY` is configured in repository secrets
+
+### API Errors
+
+If the GLM API call fails:
+- Check the Actions tab for detailed error logs
+- Verify the `GLM_API_KEY` secret is correctly configured
+- Ensure sufficient API quota is available
+- The bot will automatically retry up to 3 times with exponential backoff
+
+### Review Seems Truncated
+
+If the review appears incomplete:
+- Large diffs may be truncated at 100,000 characters
+- Check the logs for truncation warnings
+- Consider breaking large PRs into smaller chunks
+
+## Configuration
+
+### Required Secrets
+
+The following secret must be configured in the repository settings:
+
+- `GLM_API_KEY` - Your GLM (BigModel) API key for accessing the GLM-4.7 API
+
+To add the secret:
+1. Go to repository Settings → Secrets and variables → Actions
+2. Click "New repository secret"
+3. Name: `GLM_API_KEY`
+4. Value: Your GLM API key
+
+### Optional Configuration
+
+The following optional environment variables can be set in the workflow file:
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `GLM_API_URL` | `https://open.bigmodel.cn/api/paas/v4/chat/completions` | GLM API endpoint |
+| `GLM_MODEL` | `glm-4.7` | Model to use for reviews |
+| `PR_REVIEWER_COOLDOWN_MINUTES` | `5` | Cooldown period between reviews |
+| `PR_REVIEWER_MAX_RETRIES` | `3` | Maximum API retry attempts |
+| `PR_REVIEWER_RETRY_DELAY` | `1.0` | Base delay for retry backoff (seconds) |
+| `PR_REVIEWER_MAX_DIFF_SIZE` | `100000` | Maximum diff size before truncation |
+
+### Workflow Customization
+
+The workflow can be customized in `.github/workflows/pr-reviewer.yml`:
+- Change Python version (default: 3.11)
+- Adjust timeout value (default: 10 minutes)
+- Modify trigger conditions
+- Add additional dependencies
+
+## Code Quality
+
+The PR reviewer script follows vllm-omni coding standards:
+
+- **Type hints**: All functions have complete type hints following mypy strict mode
+- **Logging**: Uses Python's logging module for structured logging
+- **Testing**: Comprehensive unit tests with pytest
+- **Pre-commit**: Script is checked by pre-commit hooks (flake8)
+
+## Cost Estimate
+
+| Component | Cost |
+|-----------|------|
+| GitHub Actions (public repo) | Free |
+| GLM API (glm-4.7) | ~0.50-5 CNY per PR (varies by size) |
+| Total (20 PRs/month) | ~10-100 CNY/month (~$2-15 USD) |
+
+## Contributing
+
+To improve the PR reviewer bot:
+
+1. Edit `.github/scripts/pr_reviewer.py` for logic changes
+2. Edit `.github/workflows/pr-reviewer.yml` for workflow changes
+3. Add tests to `.github/tests/test_pr_reviewer.py`
+4. Run `pre-commit run --files .github/scripts/pr_reviewer.py` to check code quality
+5. Test thoroughly with a test PR before deploying to production
+
+## License
+
+This bot is part of the VLLM-Omni project and follows the same license terms.

docs/serving/speech_api.md

@@ -0,0 +1,267 @@
+# Speech API
+
+vLLM-Omni provides an OpenAI-compatible API for text-to-speech (TTS) generation using Qwen3-TTS models.
+
+Each server instance runs a single model (specified at startup via `vllm serve <model> --omni`).
+
+## Quick Start
+
+### Start the Server
+
+```bash
+# CustomVoice model (predefined speakers)
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts.yaml \
+    --omni \
+    --port 8091 \
+    --trust-remote-code \
+    --enforce-eager
+```
+
+### Generate Speech
+
+**Using curl:**
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "Hello, how are you?",
+        "voice": "vivian",
+        "language": "English"
+    }' --output output.wav
+```
+
+**Using Python:**
+
+```python
+import httpx
+
+response = httpx.post(
+    "http://localhost:8091/v1/audio/speech",
+    json={
+        "input": "Hello, how are you?",
+        "voice": "vivian",
+        "language": "English",
+    },
+    timeout=300.0,
+)
+
+with open("output.wav", "wb") as f:
+    f.write(response.content)
+```
+
+**Using OpenAI SDK:**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:8091/v1", api_key="none")
+
+response = client.audio.speech.create(
+    model="Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
+    voice="vivian",
+    input="Hello, how are you?",
+)
+
+response.stream_to_file("output.wav")
+```
+
+## API Reference
+
+### Endpoint
+
+```
+POST /v1/audio/speech
+Content-Type: application/json
+```
+
+### Request Parameters
+
+#### OpenAI Standard Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `input` | string | **required** | The text to synthesize into speech |
+| `model` | string | server's model | Model to use (optional, should match server if specified) |
+| `voice` | string | "vivian" | Speaker name (e.g., vivian, ryan, aiden) |
+| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, aac, opus |
+| `speed` | float | 1.0 | Playback speed (0.25-4.0) |
+
+#### vLLM-Omni Extension Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `task_type` | string | "CustomVoice" | TTS task type: CustomVoice, VoiceDesign, or Base |
+| `language` | string | "Auto" | Language (see supported languages below) |
+| `instructions` | string | "" | Voice style/emotion instructions |
+| `max_new_tokens` | integer | 2048 | Maximum tokens to generate |
+
+**Supported languages:** Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
+
+#### Voice Clone Parameters (Base task)
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `ref_audio` | string | null | Reference audio (URL or base64 data URL) |
+| `ref_text` | string | null | Transcript of reference audio |
+| `x_vector_only_mode` | bool | null | Use speaker embedding only (no ICL) |
+
+### Response Format
+
+Returns binary audio data with appropriate `Content-Type` header (e.g., `audio/wav`).
+
+### Voices Endpoint
+
+```
+GET /v1/audio/voices
+```
+
+Lists available voices for the loaded model.
+
+```json
+{
+    "voices": ["aiden", "dylan", "eric", "ono_anna", "ryan", "serena", "sohee", "uncle_fu", "vivian"]
+}
+```
+
+## Examples
+
+### CustomVoice with Style Instruction
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "I am so excited!",
+        "voice": "vivian",
+        "instructions": "Speak with great enthusiasm"
+    }' --output excited.wav
+```
+
+### VoiceDesign (Natural Language Voice Description)
+
+```bash
+# Start server with VoiceDesign model first
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts.yaml \
+    --omni \
+    --port 8091 \
+    --trust-remote-code \
+    --enforce-eager
+```
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "Hello world",
+        "task_type": "VoiceDesign",
+        "instructions": "A warm, friendly female voice with a gentle tone"
+    }' --output designed.wav
+```
+
+### Base (Voice Cloning)
+
+```bash
+# Start server with Base model first
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts.yaml \
+    --omni \
+    --port 8091 \
+    --trust-remote-code \
+    --enforce-eager
+```
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "Hello, this is a cloned voice",
+        "task_type": "Base",
+        "ref_audio": "https://example.com/reference.wav",
+        "ref_text": "Original transcript of the reference audio"
+    }' --output cloned.wav
+```
+
+## Supported Models
+
+| Model | Task Type | Description |
+|-------|-----------|-------------|
+| `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` | CustomVoice | Predefined speaker voices with optional style control |
+| `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | VoiceDesign | Natural language voice style description |
+| `Qwen/Qwen3-TTS-12Hz-1.7B-Base` | Base | Voice cloning from reference audio |
+| `Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice` | CustomVoice | Smaller/faster variant |
+| `Qwen/Qwen3-TTS-12Hz-0.6B-Base` | Base | Smaller/faster variant for voice cloning |
+
+## Error Responses
+
+### 400 Bad Request
+
+Invalid parameters:
+
+```json
+{
+    "error": {
+        "message": "Input text cannot be empty",
+        "type": "BadRequestError",
+        "param": null,
+        "code": 400
+    }
+}
+```
+
+### 404 Not Found
+
+Model not found:
+
+```json
+{
+    "error": {
+        "message": "The model `xxx` does not exist.",
+        "type": "NotFoundError",
+        "param": "model",
+        "code": 404
+    }
+}
+```
+
+## Troubleshooting
+
+### "TTS model did not produce audio output"
+
+Ensure you're using the correct model variant for your task type:
+- CustomVoice task → CustomVoice model
+- VoiceDesign task → VoiceDesign model
+- Base task → Base model
+
+### Server Not Running
+
+```bash
+# Check if server is responding
+curl http://localhost:8091/v1/audio/voices
+```
+
+### Out of Memory
+
+If you encounter OOM errors:
+1. Use smaller model variant: `Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice`
+2. Reduce `--gpu-memory-utilization`
+
+### Unsupported Speaker
+
+Use `/v1/audio/voices` to list available voices for the loaded model.
+
+## Development
+
+Enable debug logging:
+
+```bash
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts.yaml \
+    --omni \
+    --port 8091 \
+    --trust-remote-code \
+    --enforce-eager \
+    --uvicorn-log-level debug
+```

docs/user_guide/diffusion/cache_dit_acceleration.md

@@ -188,8 +188,8 @@ cd examples/offline_inference/text_to_image
 python text_to_image.py \
     --model Qwen/Qwen-Image \
     --prompt "a cup of coffee on the table" \
-    --cache_backend cache_dit \
-    --num_inference_steps 50
+    --cache-backend cache_dit \
+    --num-inference-steps 50
 ```
 
 

docs/user_guide/diffusion/cpu_offload_diffusion.md

@@ -1,101 +1,143 @@
-# CPU Offloading for Diffusion Model
+# CPU Offloading for Diffusion Models
 
 ## Overview
 
-vLLM-Omni provides two offloading strategies to reduce GPU memory usage for diffusion models, allowing you to run larger models on GPUs with limited VRAM:
+vLLM-Omni provides two offloading strategies to reduce GPU memory usage for diffusion models:
 
-1. **Model-level (Component) Offloading**: Swaps entire model components (DiT transformer, VAE, encoders) between GPU and CPU.
-2. **Layerwise (Blockwise) Offloading**: Keeps only a single or a few transformer blocks on GPU at a time, with compute - memory copy overlap.
+1. **Model-level (Sequential) Offloading**: Mutual exclusion between DiT model and encoder - only one is on GPU at a time.
+2. **Layerwise (Blockwise) Offloading**: Keeps only one transformer block on GPU at a time with compute-memory overlap.
 
-Both approaches use pinned memory for faster CPU-GPU transfers. For now, the two offloading strategies could not be used at the same time.
+Both strategies use pinned memory for faster CPU-GPU transfers. The strategies are **mutually exclusive** for now - if both are enabled, layerwise takes priority.
 
 
-## Model-level CPU Offloading
+## Model-level (Sequential) Offloading
 
-### Implementation
+### How It Works
 
-CPU offload lets the diffusion worker move large model components between GPU and CPU memory on demand. It keeps the DiT transformer resident on GPU only while it is actively running, and swaps it out when encoders modules need the device. This reduces peak VRAM usage so bigger checkpoints run on smaller GPUs, or multiple requests can share the same GPU.
+Model-level offloading implements mutual exclusion between DiT transformer and encoder modules using pre forward hooks:
 
-**Execution Flow**:
-1. Text encoders run on GPU while the DiT transformer is offloaded to CPU.
-2. Before denoising, weights are prefetched back to GPU, honoring pinned-memory copies for speed.
-3. After the diffusion step, the transformer returns to CPU and the process repeats as needed.
+- **When encoders run**: DiT transformer is offloaded to CPU
+- **When DiT runs**: Encoders are offloaded to CPU
+- **VAE**: Stays resident on GPU
 
-Transfers use pinned host buffers, and the worker coordinates swaps via mutex-style hooks so components never compete for memory.
+Before each module's forward pass, the hook automatically moves it to GPU while offloading the other module group to CPU. Transfers use pinned memory for speed.
 
-### Configuration
-You can enable CPU offload in two ways:
-
-1. **Python API**: set `enable_cpu_offload=True`.
+### Usage
 
+**Python API:**
 ```python
 from vllm_omni import Omni
 
-if __name__ == "__main__":
-
-    m = Omni(model="Qwen/Qwen-Image",enable_cpu_offload=True)
+m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_cpu_offload=True)
 ```
 
-2. **CLI**: pass `--enable-cpu-offload` to the diffusion service entrypoint.
+**CLI:**
+```bash
+vllm-omni serve diffusion Wan-AI/Wan2.2-T2V-A14B-Diffusers --enable-cpu-offload
+```
 
 ### Limitations
-- Cold start latency increases for over one minute for some models(e.g., Qwen-Image)
+- Cold start latency increases
+- Adds overhead from CPU-GPU transfers between encoder and denoising phases
+- Support single GPU only for now
 
 
 ## Layerwise (Blockwise) Offloading
 
-### Implementation
-Layerwise offload operates at transformer block granularity, keeping a single transformer block, or a specified number of blocks, on GPU while others stay in CPU memory.
+### How It Works
 
-Unlike full model-wise CPU offload which swaps entire components like DiT and encoders, layerwise offloading applies a sliding window way of loading and offloading weights between gpu and cpu: while block `i` computes, block `i+1` get prefetched asynchronously via pinned memory. In this way, only partial blocks(s) reside on GPU at any moment during inference, so that greatly decrease the memory occupancy.
+Layerwise offloading keeps only one transformer block on GPU at a time.
 
-**Execution Flow**:
+As each block completes, the next block is prefetched to GPU while the current block is freed. The pre and forward hooks utilized by layerwise offloading apply a separate CUDA stream (`copy_stream`) to overlap weight transfer with computation, and retain flattened tensors in pinned CPU memory for block parameters re-materialization. Encoders, VAE, and non-block DiT modules (embeddings, norms) always stay on GPU.
 
-1. During model initialization, all components are loaded to CPU first. Then components other than DiT model(s) in the pipeline, such as VAE and encoders, are moved to GPU. The weights of target transformer blocks are collected as contiguous tensors per layer on CPU with pinned memory; and non-block modules (embeddings, norms, etc) in the DiT model are moved to and stay on GPU.
-2. The first block(s) are transferred to GPU during initialization of `LayerwiseOffloader`, before the first denoising step of the very first request.
-3. As each block executes, the next block prefetches on a separate CUDA stream for compute - memory copy overlap. After execution, the current block is immediately freed from GPU memory.
-4. When the last block completes, the first block prefetches for the next denoising step.
+**Execution Flow:**
 
+| Block | Pre-forward Hook | Forward | Post-forward Hook |
+|-------|------------------|---------|-------------------|
+| block-0 | Prefetch block-1 (async) | Compute block-0 | Free block-0 |
+| block-1 | Prefetch block-2 (async) | Compute block-1 | Free block-1 |
+| ... | ... | ... | ... |
+| block-(n-1) | **Prefetch block-0** (async) | Compute block-(n-1) | Free block-(n-1) |
 
-Example of hook executions of a DiT model with n layers, by default keep a single layer on GPU:
-| Layer (block) idx | forward pre-hook               | forward          | forward post-hook         |
-|-------------------|--------------------------------|------------------|---------------------------|
-| layer-0           | prefetch layer 1 (copy stream) | compute layer 0  | free layer-0 gpu weights  |
-| layer-1           | prefetch layer 2 (copy stream) | compute layer 1  | free layer-1 gpu weights  |
-| layer-2           | prefetch layer 3 (copy stream) | compute layer 2  | free layer-2 gpu weights  |
-| ...               | ...                            | ...              | ...                       |
-| layer-(n-1)       | **prefetch layer 0 (copy stream)** | compute layer (n-1) | free layer (n-1) gpu weights  |
+Each transformer block has a `LayerwiseOffloadHook` that prefetches the next block before forward and frees the current block after forward.
 
+Layerwise offloading is primarily recommended for large **video generation models** where the compute cost per block is high enough to effectively overlap with memory prefetch operations. For example, Wan2.2 T2V and I2V pipelines.
 
-### Configuration
-
-1. **Python API**: set `enable_layerwise_offload=True` and optionally `layerwise_num_gpu_layers`.
+### Usage
 
+**Python API:**
 ```python
 from vllm_omni import Omni
 
-if __name__ == "__main__":
-    m = Omni(
-        model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
-        enable_layerwise_offload=True,
-        ...
-    )
+# Text-to-video
+m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_layerwise_offload=True)
+
+# Or image-to-video
+m = Omni(model="Wan-AI/Wan2.2-I2V-A14B-Diffusers", enable_layerwise_offload=True)
 ```
 
-2. **CLI**: pass `--enable-layerwise-offload` and `--layerwise-num-gpu-layers` to the diffusion service entrypoint.
+**CLI:**
+```bash
+# Text-to-video
+vllm-omni serve diffusion Wan-AI/Wan2.2-T2V-A14B-Diffusers --enable-layerwise-offload
 
-### Supported Models
+# Or image-to-video
+vllm-omni serve diffusion Wan-AI/Wan2.2-I2V-A14B-Diffusers --enable-layerwise-offload
+```
 
-| Architecture | Models | Example HF Models | DiT Model Cls | Blocks Attr Name |
-|--------------|--------|-------------------|----------|----------|
-| `QwenImagePipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | "transformer_blocks" |
-| `Wan22Pipeline` | Wan2.2 | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | "blocks" |
+### To Support a Model
 
-NOTE: Models must define `_layerwise_offload_blocks_attr` class attribute so that the layerwise offloader finds the target transformer blocks.
+Models must define the blocks attribute name for layerwise offloading:
+
+```python
+class WanTransformer3DModel(nn.Module):
+    _layerwise_offload_blocks_attr = "blocks"  # Attribute name containing transformer blocks
+
+    def __init__(self):
+        self.blocks = nn.ModuleList([...])  # Transformer blocks
+```
 
 ### Limitations
 - Cold start latency increases because of
-    1) components are loaded to CPU first at the very first during initialization,  
+    1) components are loaded to CPU first at the very first during initialization,
     2) weight consolidation and pinning
-- Performance depends on CPU <-> GPU interconnection (e.g., PCIe bandwidth).
+- Performance depends on compute cost and H2D bandwidth as well
 - Support single GPU only for now
+
+
+### Implementation Notes
+
+**Module Discovery**
+
+The offloader automatically discovers pipeline components:
+
+- **DiT modules**: `transformer`, `transformer_2`, `dit`
+- **Encoders**: `text_encoder`, `text_encoder_2`, `text_encoder_3`, `image_encoder`
+- **VAE**: `vae`
+
+**Hook System**
+
+Both strategies use vLLM-Omni's hook registry system (`HookRegistry` and `ModelHook`) to register pre/post forward callbacks on modules, enabling automatic swapping without modifying model code.
+
+**Backend Architecture**
+
+```
+OffloadBackend (base class)
+├── ModelLevelOffloadBackend → uses SequentialOffloadHook
+└── LayerWiseOffloadBackend → uses LayerwiseOffloadHook
+```
+
+Factory function `get_offload_backend()` selects the appropriate backend based on configuration.
+
+
+## Supported Models
+
+| Architecture | Example Models | DiT Class | Model-Level Offload | Layerwise Offload | Blocks Attr (Layerwise specific) |
+|--------------|----------------|-----------|---------------------|-------------------|-------------|
+| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | `"blocks"` |
+| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | `"blocks"` |
+| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | ✓ | ✓ | `"transformer_blocks"` |
+
+**Notes:**
+- Model-Level Offloading is expected to be supported by all common diffusion models (DiT and encoders) naturally
+- Layerwise Offloading requires DiT class to define `_layerwise_offload_blocks_attr` pointing to transformer blocks

docs/user_guide/diffusion/lora.md

@@ -0,0 +1,63 @@
+# LoRA (Low-Rank Adaptation) Guide
+
+LoRA (Low-Rank Adaptation) enables fine-tuning diffusion models by adding trainable low-rank matrices to existing model weights. vLLM-Omni currently supports PEFT-style LoRA adapters, allowing you to customize model behavior without modifying the base model weights.
+
+## Overview
+
+LoRA adapters are lightweight, model-specific fine-tuning weights that can be dynamically loaded and applied to diffusion models. vLLM-Omni uses a unified LoRA handling mechanism similar to vLLM with LRU cache management.
+
+## LoRA Adapter Format
+
+LoRA adapters must be in **PEFT (Parameter-Efficient Fine-Tuning)** format. A typical LoRA adapter directory structure:
+
+```
+lora_adapter/
+├── adapter_config.json
+└── adapter_model.safetensors
+```
+
+The `adapter_config.json` file contains metadata about the LoRA adapter, including:
+- `r`: LoRA rank
+- `lora_alpha`: LoRA alpha scaling factor
+- `target_modules`: List of module names to apply LoRA to
+
+## Quick Start
+
+### Offline Inference
+
+#### Pre-loaded LoRA
+
+Load a LoRA adapter at initialization. This adapter is pre-loaded into the cache and can be activated by requests:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.lora.request import LoRARequest
+
+lora_path="/path/to/lora_adapter"
+
+omni = Omni(
+    model="stabilityai/stable-diffusion-3.5-medium",
+    lora_path=lora_path
+)
+
+lora_request = LoRARequest(
+    lora_name="preloaded",
+    lora_int_id=1,
+    lora_path=lora_path
+)
+
+outputs = omni.generate(
+    prompt="A piece of cheesecake",
+    lora_request=lora_request,
+    lora_scale=2.0, # optional arg, default 1.0
+)
+```
+
+!!! note "Server-side Path Requirement"
+    The LoRA adapter path (`local_path`) must be readable on the **server** machine. If your client and server are on different machines, ensure the LoRA adapter is accessible via a shared mount or copied to the server.
+
+
+## See Also
+
+- [Text-to-Image Offline Example](../examples/offline_inference/text_to_image.md#lora) - Complete offline LoRA example
+- [Text-to-Image Online Example](../examples/online_serving/text_to_image.md#lora) - Complete online LoRA example

docs/user_guide/diffusion/parallelism_acceleration.md

@@ -14,23 +14,27 @@ The following parallelism methods are currently supported in vLLM-Omni:
 
 4. [Tensor Parallelism](#tensor-parallelism): Tensor parallelism shards model weights across devices. This can reduce per-GPU memory usage. Note that for diffusion models we currently shard the majority of layers within the DiT.
 
+5. [VAE Patch Parallelism](#vae-patch-parallelism): VAE patch parallelism shards VAE decode spatially across ranks. This can reduce the peak memory of VAE decode and (depending on resolution and communication overhead) speed up VAE decode.
+
+6. [HSDP](#hsdp): Hybrid Sharded Data Parallel shards model weights across GPUs using PyTorch FSDP2. This reduces per-GPU memory usage, enabling inference of large models on GPUs with limited memory.
+
 The following table shows which models are currently supported by parallelism method:
 
 ### ImageGen
 
-| Model                    | Model Identifier                     | Ulysses-SP | Ring-SP | CFG-Parallel | Tensor-Parallel |
-|--------------------------|--------------------------------------|:----------:|:-------:|:------------:|:---------------:|
-| **LongCat-Image**        | `meituan-longcat/LongCat-Image`      |     ✅      |    ✅    |      ❌       |        ✅        |
-| **LongCat-Image-Edit**   | `meituan-longcat/LongCat-Image-Edit` |     ✅      |    ✅    |      ❌       |        ✅        |
-| **Ovis-Image**           | `OvisAI/Ovis-Image`                  |     ❌      |    ❌    |      ❌       |        ❌        |
-| **Qwen-Image**           | `Qwen/Qwen-Image`                    |     ✅      |    ✅    |      ✅       |        ✅        |
-| **Qwen-Image-Edit**      | `Qwen/Qwen-Image-Edit`               |     ✅      |    ✅    |      ✅       |        ✅        |
-| **Qwen-Image-Edit-2509** | `Qwen/Qwen-Image-Edit-2509`          |     ✅      |    ✅    |      ✅       |        ✅        |
-| **Qwen-Image-Layered**   | `Qwen/Qwen-Image-Layered`            |     ✅      |    ✅    |      ✅       |        ✅        |
-| **Z-Image**              | `Tongyi-MAI/Z-Image-Turbo`           |     ✅      |    ✅    |      ❌       |  ✅ (TP=2 only)  |
-| **Stable-Diffusion3.5**  | `stabilityai/stable-diffusion-3.5`   |     ❌      |    ❌    |      ❌       |        ❌        |
-| **FLUX.2-klein**         | `black-forest-labs/FLUX.2-klein-4B`  |     ❌      |    ❌    |      ❌       |        ✅        |
-| **FLUX.1-dev**           | `black-forest-labs/FLUX.1-dev`       |     ❌      |    ❌    |      ❌       |        ✅        |
+| Model                    | Model Identifier                     | Ulysses-SP | Ring-SP | CFG-Parallel | Tensor-Parallel | VAE-Patch-Parallel |
+|--------------------------|--------------------------------------|:----------:|:-------:|:------------:|:---------------:|:------------------:|
+| **LongCat-Image**        | `meituan-longcat/LongCat-Image`      |     ✅      |    ✅    |      ❌       |        ✅        |         ❌          |
+| **LongCat-Image-Edit**   | `meituan-longcat/LongCat-Image-Edit` |     ✅      |    ✅    |      ❌       |        ✅        |         ❌          |
+| **Ovis-Image**           | `OvisAI/Ovis-Image`                  |     ❌      |    ❌    |      ❌       |        ❌        |         ❌          |
+| **Qwen-Image**           | `Qwen/Qwen-Image`                    |     ✅      |    ✅    |      ✅       |        ✅        |         ❌          |
+| **Qwen-Image-Edit**      | `Qwen/Qwen-Image-Edit`               |     ✅      |    ✅    |      ✅       |        ✅        |         ❌          |
+| **Qwen-Image-Edit-2509** | `Qwen/Qwen-Image-Edit-2509`          |     ✅      |    ✅    |      ✅       |        ✅        |         ❌          |
+| **Qwen-Image-Layered**   | `Qwen/Qwen-Image-Layered`            |     ✅      |    ✅    |      ✅       |        ✅        |         ❌          |
+| **Z-Image**              | `Tongyi-MAI/Z-Image-Turbo`           |     ✅      |    ✅    |      ❌       |  ✅ (TP=2 only)  |         ✅          |
+| **Stable-Diffusion3.5**  | `stabilityai/stable-diffusion-3.5`   |     ❌      |    ❌    |      ❌       |        ✅        |         ✅          |
+| **FLUX.2-klein**         | `black-forest-labs/FLUX.2-klein-4B`  |     ❌      |    ❌    |      ❌       |        ✅        |         ❌          |
+| **FLUX.1-dev**           | `black-forest-labs/FLUX.1-dev`       |     ❌      |    ❌    |      ✅       |        ✅        |         ❌          |
 
 !!! note "TP Limitations for Diffusion Models"
     We currently implement Tensor Parallelism (TP) only for the DiT (Diffusion Transformer) blocks. This is because the `text_encoder` component in vLLM-Omni uses the original Transformers implementation, which does not yet support TP.
@@ -47,9 +51,9 @@ The following table shows which models are currently supported by parallelism me
 
 ### VideoGen
 
-| Model | Model Identifier | Ulysses-SP | Ring-SP | Tensor-Parallel |
-|-------|------------------|------------|---------|--------------------------|
-| **Wan2.2** | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | ✅ | ✅ | ✅ |
+| Model | Model Identifier | Ulysses-SP | Ring-Attention | Tensor-Parallel | HSDP |
+|-------|------------------|:----------:|:--------------:|:---------------:|:----:|
+| **Wan2.2** | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | ✅ | ✅ | ✅ | ✅ |
 
 ### Tensor Parallelism
 
@@ -76,6 +80,45 @@ outputs = omni.generate(
 )
 ```
 
+### VAE Patch Parallelism
+
+VAE patch parallelism distributes the VAE decode workload across multiple ranks by splitting the latent spatially. It is configured via `DiffusionParallelConfig.vae_patch_parallel_size` and can be combined with other parallelism methods (e.g., TP).
+
+!!! note "Enablement and feature gate"
+    - VAE patch parallelism is currently **enabled only for validated pipelines** (currently: `Tongyi-MAI/Z-Image-Turbo`, `stabilityai/stable-diffusion-3.5`).
+    - If `vae_patch_parallel_size > 1` is set for a validated pipeline, vLLM-Omni will automatically enable `vae_use_tiling` as a safety gate. (We use `vae_use_tiling` because it indicates the VAE supports diffusers tiling parameters like `tile_latent_min_size` and `tile_overlap_factor`.)
+
+#### Offline Inference
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="Tongyi-MAI/Z-Image-Turbo",
+    parallel_config=DiffusionParallelConfig(
+        tensor_parallel_size=2,
+        vae_patch_parallel_size=2,
+    ),
+    vae_use_tiling=True,
+)
+
+outputs = omni.generate(
+    prompt="a cat reading a book",
+    num_inference_steps=9,
+    width=1024,
+    height=1024,
+)
+```
+
+#### How it works (method selection)
+
+VAE patch parallelism automatically selects between two internal decode methods based on whether diffusers tiling would kick in:
+
+- `_distributed_tiled_decode`: Used when the latent spatial size exceeds `vae.tile_latent_min_size` (i.e., diffusers tiled decode). Each rank decodes a subset of tiles; rank0 gathers and runs the same overlap+blend+stitch logic as diffusers. This matches the single-rank diffusers tiled output.
+
+- `_distributed_patch_decode`: Used when diffusers tiling would not kick in. Each rank decodes a grid patch expanded with a latent-space halo; then rank0 gathers the cropped core patches and stitches them into the full image. This path has no blending and can introduce small numerical differences compared to the non-parallel decode.
+
 ### Sequence Parallelism
 
 #### Ulysses-SP
@@ -233,143 +276,6 @@ To measure the parallelism methods, we run benchmarks with **Qwen/Qwen-Image** m
 | Hybrid Ulysses + Ring  |  2  |  2  |  24.3s | 1.87x |
 
 
-##### How to parallelize a new model
-
-NOTE: "Terminology: SP vs CP"
-    Our "Sequence Parallelism" (SP) corresponds to "Context Parallelism" (CP) in the [diffusers library](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/_modeling_parallel.py).
-    We use "Sequence Parallelism" to align with vLLM-Omni's terminology.
-
----
-
-###### Non-intrusive `_sp_plan` (Recommended)
-
-The `_sp_plan` mechanism allows SP without modifying `forward()` logic. The framework automatically registers hooks to shard inputs and gather outputs at module boundaries.
-
-**Requirements for `forward()` function:**
-
-- Tensor operations that need sharding/gathering must happen at **`nn.Module` boundaries** (not inline Python operations)
-- If your `forward()` contains inline tensor operations (e.g., `torch.cat`, `pad_sequence`) that need sharding, **extract them into a submodule**
-
-**When to create a submodule:**
-
-```python
-# ❌ BAD: Inline operations - hooks cannot intercept
-def forward(self, x, cap_feats):
-    unified = torch.cat([x, cap_feats], dim=1)  # Cannot be sharded via _sp_plan
-    ...
-
-# ✅ GOOD: Extract into a submodule
-class UnifiedPrepare(nn.Module):
-    def forward(self, x, cap_feats):
-        return torch.cat([x, cap_feats], dim=1)  # Now can be sharded via _sp_plan
-
-class MyModel(nn.Module):
-    def __init__(self):
-        self.unified_prepare = UnifiedPrepare()  # Submodule
-
-    def forward(self, x, cap_feats):
-        unified = self.unified_prepare(x, cap_feats)  # Hook can intercept here
-```
-
----
-
-###### Defining `_sp_plan`
-
-**Type definitions** (see [diffusers `_modeling_parallel.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/_modeling_parallel.py) for reference):
-
-```python
-from vllm_omni.diffusion.distributed.sp_plan import (
-    SequenceParallelInput,   # Corresponds to diffusers' ContextParallelInput
-    SequenceParallelOutput,  # Corresponds to diffusers' ContextParallelOutput
-)
-```
-
-| Parameter | Description |
-|-----------|-------------|
-| `split_dim` | Dimension to split/gather (usually `1` for sequence) |
-| `expected_dims` | Expected tensor rank for validation (optional) |
-| `split_output` | `False`: shard **input** parameters; `True`: shard **output** tensors |
-| `auto_pad` | Auto-pad if sequence not divisible by world_size (Ulysses only) |
-
-**Key naming convention:**
-
-| Key | Meaning | Python equivalent |
-|-----|---------|-------------------|
-| `""` | Root model | `model` |
-| `"blocks.0"` | First element of ModuleList | `model.blocks[0]` |
-| `"blocks.*"` | All elements of ModuleList | `for b in model.blocks` |
-| `"outputs.main"` | ModuleDict entry | `model.outputs["main"]` |
-
-**Dictionary key types:**
-
-| Key type | `split_output` | Description |
-|----------|----------------|-------------|
-| `"param_name"` (str) | `False` | Shard **input parameter** by name |
-| `0`, `1` (int) | `True` | Shard **output tuple** by index |
-
-**Example** (similar to [diffusers `transformer_wan.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_wan.py)):
-
-```python
-class MyTransformer(nn.Module):
-    _sp_plan = {
-        # Shard rope module OUTPUTS (returns tuple)
-        "rope": {
-            0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),  # cos
-            1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),  # sin
-        },
-        # Shard transformer block INPUT parameter
-        "blocks.0": {
-            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
-        },
-        # Gather at final projection
-        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
-    }
-```
-
----
-
-###### Hook flow
-
-```
-Input → [SequenceParallelSplitHook: pre_forward] → Module.forward() → [post_forward] → ...
-                                                                              ↓
-... → [SequenceParallelGatherHook: post_forward] → Output
-```
-
-1. **SplitHook** shards tensors before/after the target module
-2. **Attention layers** handle Ulysses/Ring communication internally
-3. **GatherHook** collects sharded outputs
-
-The framework automatically applies these hooks when `sequence_parallel_size > 1`.
-
----
-
-###### Method 2: Intrusive modification (For complex cases)
-
-For models with dynamic sharding logic that cannot be expressed via `_sp_plan`:
-
-```python
-from vllm_omni.diffusion.distributed.sp_sharding import sp_shard, sp_gather
-
-def forward(self, hidden_states, ...):
-    if self.parallel_config.sequence_parallel_size > 1:
-        hidden_states = sp_shard(hidden_states, dim=1)
-        # ... computation ...
-        output = sp_gather(output, dim=1)
-    return output
-```
-
----
-
-###### Choosing the right approach
-
-| Scenario | Approach |
-|----------|----------|
-| Standard transformer | `_sp_plan` |
-| Inline tensor ops need sharding | Extract to submodule + `_sp_plan` |
-| Dynamic/conditional sharding | Intrusive modification |
-
-
 ### CFG-Parallel
 
 #### Offline Inference
@@ -413,10 +319,10 @@ python image_edit.py \
   --model "Qwen/Qwen-Image-Edit" \
   --image "qwen_image_output.png" \
   --prompt "turn this cat to a dog" \
-  --negative_prompt "low quality, blurry" \
-  --cfg_scale 4.0 \
+  --negative-prompt "low quality, blurry" \
+  --cfg-scale 4.0 \
   --output "edited_image.png" \
-  --cfg_parallel_size 2
+  --cfg-parallel-size 2
 ```
 
 #### Online Serving
@@ -427,141 +333,83 @@ You can enable CFG-Parallel in online serving for diffusion models via `--cfg-pa
 vllm serve Qwen/Qwen-Image-Edit --omni --port 8091 --cfg-parallel-size 2
 ```
 
-#### How to parallelize a pipeline
+### HSDP
 
-This section describes how to add CFG-Parallel to a diffusion **pipeline**. We use the Qwen-Image pipeline (`vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image.py`) as the reference implementation.
+HSDP (Hybrid Sharded Data Parallel) shards model weights across GPUs to reduce per-GPU memory usage. This enables inference of large models (e.g., Wan2.2 14B) on GPUs with limited memory.
 
-In `QwenImagePipeline`, each diffusion step runs two denoiser forward passes sequentially:
+Unlike Tensor Parallelism which splits computation, HSDP uses PyTorch's FSDP2 to shard and redistribute weights at runtime. Each GPU only holds a fraction of the model weights, and weights are gathered on-demand during forward passes.
 
-- positive (prompt-conditioned)
-- negative (negative-prompt-conditioned)
+#### Configuration
 
-CFG-Parallel assigns these two branches to different ranks in the **CFG group** and synchronizes the results.
+HSDP is configured via `DiffusionParallelConfig`:
 
-vLLM-omni provides `CFGParallelMixin` base class that encapsulates the CFG parallel logic. By inheriting from this mixin and calling its methods, pipelines can easily implement CFG parallel without writing repetitive code.
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `use_hsdp` | bool | False | Enable HSDP |
+| `hsdp_shard_size` | int | -1 | Number of GPUs to shard weights across. -1 = auto (requires other parallelism > 1) |
+| `hsdp_replicate_size` | int | 1 | Number of replica groups. Each group holds a full sharded copy |
 
-**Key Methods in CFGParallelMixin:**
-- `predict_noise_maybe_with_cfg()`: Automatically handles CFG parallel noise prediction
-- `scheduler_step_maybe_with_cfg()`: Scheduler step with automatic CFG rank synchronization
+**Constraints:**
 
-**Example Implementation:**
+- `hsdp_replicate_size × hsdp_shard_size == world_size`
+- HSDP cannot be used with Tensor Parallelism (`tensor_parallel_size` must be 1)
+
+#### Operating Modes
+
+HSDP can work in two modes:
+
+- **Standalone Mode**: HSDP alone without other parallelism. Must specify `hsdp_shard_size` explicitly.
+- **Combined Mode**: HSDP overlays on top of other parallelism (Ulysses Sequence Parallel, CFG Parallel). HSDP dimensions must match world_size.
+
+#### Offline Inference
+
+**Standalone HSDP** (shard across 4 GPUs, no other parallelism):
 
 ```python
-class QwenImageCFGParallelMixin(CFGParallelMixin):
-    """
-    Base Mixin class for Qwen Image pipelines providing shared CFG methods.
-    """
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
 
-    def diffuse(
-        self,
-        prompt_embeds: torch.Tensor,
-        prompt_embeds_mask: torch.Tensor,
-        negative_prompt_embeds: torch.Tensor,
-        negative_prompt_embeds_mask: torch.Tensor,
-        latents: torch.Tensor,
-        img_shapes: torch.Tensor,
-        txt_seq_lens: torch.Tensor,
-        negative_txt_seq_lens: torch.Tensor,
-        timesteps: torch.Tensor,
-        do_true_cfg: bool,
-        guidance: torch.Tensor,
-        true_cfg_scale: float,
-        image_latents: torch.Tensor | None = None,
-        cfg_normalize: bool = True,
-        additional_transformer_kwargs: dict[str, Any] | None = None,
-    ) -> torch.Tensor:
-        self.transformer.do_true_cfg = do_true_cfg
+omni = Omni(
+    model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+    parallel_config=DiffusionParallelConfig(
+        use_hsdp=True,
+        hsdp_shard_size=4,  # Shard across 4 GPUs
+    ),
+)
 
-        for i, t in enumerate(timesteps):
-            timestep = t.expand(latents.shape[0]).to(device=latents.device, dtype=latents.dtype)
-
-            # Prepare kwargs for positive (conditional) prediction
-            positive_kwargs = {
-                "hidden_states": latents,
-                "timestep": timestep / 1000,
-                "guidance": guidance,
-                "encoder_hidden_states_mask": prompt_embeds_mask,
-                "encoder_hidden_states": prompt_embeds,
-                "img_shapes": img_shapes,
-                "txt_seq_lens": txt_seq_lens,
-            }
-
-            # Prepare kwargs for negative (unconditional) prediction
-            if do_true_cfg:
-                negative_kwargs = {
-                    "hidden_states": latents,
-                    "timestep": timestep / 1000,
-                    "guidance": guidance,
-                    "encoder_hidden_states_mask": negative_prompt_embeds_mask,
-                    "encoder_hidden_states": negative_prompt_embeds,
-                    "img_shapes": img_shapes,
-                    "txt_seq_lens": negative_txt_seq_lens,
-                }
-            else:
-                negative_kwargs = None
-
-            # Predict noise with automatic CFG parallel handling
-            # - In CFG parallel mode: rank0 computes positive, rank1 computes negative
-            # - Automatically gathers results and combines them on rank0
-            noise_pred = self.predict_noise_maybe_with_cfg(
-                do_true_cfg=do_true_cfg,
-                true_cfg_scale=true_cfg_scale,
-                positive_kwargs=positive_kwargs,
-                negative_kwargs=negative_kwargs,
-                cfg_normalize=cfg_normalize,
-            )
-
-            # Step scheduler with automatic CFG synchronization
-            # - Only rank0 computes the scheduler step
-            # - Automatically broadcasts updated latents to all ranks
-            latents = self.scheduler_step_maybe_with_cfg(
-                noise_pred, t, latents, do_true_cfg
-            )
-
-        return latents
+outputs = omni.generate(
+    "A cat playing piano",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
 ```
 
-**How it works:**
-1. Prepare separate `positive_kwargs` and `negative_kwargs` for conditional and unconditional predictions
-2. Call `predict_noise_maybe_with_cfg()` which:
-   - Detects if CFG parallel is enabled (`get_classifier_free_guidance_world_size() > 1`)
-   - Distributes computation: rank0 processes positive, rank1 processes negative
-   - Gathers predictions and combines them using `combine_cfg_noise()` on rank0
-   - Returns combined noise prediction (only valid on rank0)
-3. Call `scheduler_step_maybe_with_cfg()` which:
-   - Only rank0 computes the scheduler step
-   - Broadcasts the updated latents to all ranks for synchronization
+**Combined HSDP + Sequence Parallel**:
 
-**How to customize**
-
-Some pipelines may need to customize the following functions in `CFGParallelMixin`:
-1. You may need to edit `predict_noise` function for custom behaviors.
 ```python
-def predict_noise(self, *args, **kwargs):
-    """
-    Forward pass through transformer to predict noise.
-
-    Subclasses should override this if they need custom behavior,
-    but the default implementation calls self.transformer.
-    """
-    return self.transformer(*args, **kwargs)[0]
-
+omni = Omni(
+    model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+    parallel_config=DiffusionParallelConfig(
+        ulysses_degree=4,  # Sequence parallel
+        use_hsdp=True,     # HSDP overlays on SP
+    ),
+)
 ```
-2. The default normalization function after combining the noise predictions from both branches is as follows. You may need to customize it.
-```python
-def cfg_normalize_function(self, noise_pred, comb_pred):
-    """
-    Normalize the combined noise prediction.
 
-    Args:
-        noise_pred: positive noise prediction
-        comb_pred: combined noise prediction after CFG
+#### Online Serving
 
-    Returns:
-        Normalized noise prediction tensor
-    """
-    cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
-    noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
-    noise_pred = comb_pred * (cond_norm / noise_norm)
-    return noise_pred
+**Standalone HSDP** (shard model across 4 GPUs):
+
+```bash
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --port 8091 --use-hsdp --hsdp-shard-size 4
 ```
+
+**Combined with Sequence Parallel**:
+
+```bash
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --port 8091 --use-hsdp --usp 4
+```
+
+#### Adding HSDP Support to New Models
+
+For detailed instructions on adding HSDP support to new models, see the [HSDP Contributing Guide](../../design/feature/hsdp.md).

docs/user_guide/diffusion/quantization/fp8.md

@@ -0,0 +1,77 @@
+# FP8 Quantization
+
+## Overview
+
+FP8 quantization converts BF16/FP16 weights to FP8 at model load time. No calibration or pre-quantized checkpoint needed.
+
+Depending on the model, either all layers can be quantized, or some sensitive layers should stay in BF16. See the [per-model table](#supported-models) for which case applies.
+
+Common sensitive layers in DiT-based diffusion models include **image-stream MLPs** (`img_mlp`). These are particularly vulnerable to FP8 precision loss because they process denoising latents whose dynamic range shifts significantly across timesteps, and unlike attention projections (which benefit from QK-Norm stabilization), MLPs have no built-in normalization to absorb quantization error. In deep architectures (e.g., 60+ residual blocks), small per-layer errors compound and degrade output quality. Other layers such as **attention projections** (`to_qkv`, `to_out`) and **text-stream MLPs** (`txt_mlp`) are generally more robust due to normalization or more stable input statistics.
+
+## Configuration
+
+1. **Python API**: set `quantization="fp8"`. To skip sensitive layers, use `quantization_config` with `ignored_layers`.
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+# All layers quantized
+omni = Omni(model="<your-model>", quantization="fp8")
+
+# Skip sensitive layers
+omni = Omni(
+    model="<your-model>",
+    quantization_config={
+        "method": "fp8",
+        "ignored_layers": ["<layer-name>"],
+    },
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+2. **CLI**: pass `--quantization fp8` and optionally `--ignored-layers`.
+
+```bash
+# All layers
+python text_to_image.py --model <your-model> --quantization fp8
+
+# Skip sensitive layers
+python text_to_image.py --model <your-model> --quantization fp8 --ignored-layers "img_mlp"
+
+# Online serving
+vllm serve <your-model> --omni --quantization fp8
+```
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `method` | str | — | Quantization method (`"fp8"`) |
+| `ignored_layers` | list[str] | `[]` | Layer name patterns to keep in BF16 |
+| `activation_scheme` | str | `"dynamic"` | `"dynamic"` (no calibration) or `"static"` |
+| `weight_block_size` | list[int] \| None | `None` | Block size for block-wise weight quantization |
+
+The available `ignored_layers` names depend on the model architecture (e.g., `to_qkv`, `to_out`, `img_mlp`, `txt_mlp`). Consult the transformer source for your target model.
+
+## Supported Models
+
+| Model | HF Models | Recommendation | `ignored_layers` |
+|-------|-----------|---------------|------------------|
+| Z-Image | `Tongyi-MAI/Z-Image-Turbo` | All layers | None |
+| Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | Skip sensitive layers | `img_mlp` |
+
+## Combining with Other Features
+
+FP8 quantization can be combined with cache acceleration:
+
+```python
+omni = Omni(
+    model="<your-model>",
+    quantization="fp8",
+    cache_backend="tea_cache",
+    cache_config={"rel_l1_thresh": 0.2},
+)
+```

docs/user_guide/diffusion/quantization/gguf.md

@@ -0,0 +1,185 @@
+# GGUF Quantization
+
+## Goals
+1. Reuse vLLM quantization configs and weight loaders as much as possible.
+2. Add native GGUF support to diffusion transformers without changing model definitions.
+3. Keep user-facing knobs minimal and consistent across offline and online flows.
+
+## Scope
+1. Models: Z-Image, and Flux2-klein.
+2. Components: diffusion transformer weights, loader paths, and quantization configs.
+3. Modes: native GGUF (transformer-only weights).
+
+## Architecture Overview
+1. `OmniDiffusionConfig` accepts `quantization` or `quantization_config`.
+2. Diffusion quantization wrapper (`DiffusionGgufConfig`) produces vLLM `QuantizationConfig` objects for linear layers.
+3. `DiffusersPipelineLoader` branches on quantization method and loads either HF weights or GGUF weights for the transformer.
+4. GGUF transformer loading is routed through model-specific adapters (e.g., Flux2Klein).
+5. vLLM GGUF path uses `GGUFConfig` and `GGUFLinearMethod` for matmul.
+
+## Call Chain (Offline)
+```
+CLI (examples/offline_inference/text_to_image/text_to_image.py)
+  |
+  v
+Omni (vllm_omni/entrypoints/omni.py)
+  |
+  v
+OmniStage (diffusion)
+  |
+  v
+DiffusionWorker
+  |
+  v
+DiffusionModelRunner
+  |
+  v
+DiffusersPipelineLoader
+  |
+  v
+Pipeline.forward (Flux2/Qwen/Z-Image)
+  |
+  v
+DiffusionEngine
+  |
+  v
+OmniRequestOutput
+  |
+  v
+Client (saved PNG)
+```
+
+## Call Chain (Online)
+```
+Client
+  |
+  | POST /v1/images/generations
+  v
+APIServer (vllm_omni/entrypoints/openai/api_server.py)
+  |
+  v
+_generate_with_async_omni
+  |
+  v
+AsyncOmni
+  |
+  v
+DiffusionEngine
+  |
+  v
+OmniRequestOutput
+  |
+  v
+encode_image_base64
+  |
+  v
+ImageGenerationResponse
+  |
+  v
+Client
+```
+
+## Call Chain (GGUF Operator Path)
+```
+Pipeline.forward (Flux2/Qwen/Z-Image)
+  |
+  v
+Transformer blocks
+  |
+  v
+QKVParallelLinear / ColumnParallelLinear / RowParallelLinear
+  |
+  v
+LinearBase.forward
+  |
+  v
+QuantMethod.apply (GGUFLinearMethod.apply)
+  |
+  v
+fused_mul_mat_gguf
+  |
+  v
+_fused_mul_mat_gguf (custom op)
+  |
+  v
+ops.ggml_dequantize
+  |
+  v
+x @ weight.T
+```
+
+## GGUF Weight Loading Path (Transformer-Only)
+1. `DiffusersPipelineLoader.load_model` detects `quantization_config.method == "gguf"`.
+2. `gguf_model` is resolved as one of: local file, `repo/file.gguf`, or `repo:quant_type`.
+3. GGUF weights are routed through adapters in `vllm_omni/diffusion/model_loader/gguf_adapters/`.
+4. Name mapping is applied per-architecture (Z-Image, Flux2Klein).
+5. GGUF weights are loaded into transformer modules, remaining non-transformer weights come from the HF checkpoint.
+
+## GGUF Adapter Design
+1. `GGUFAdapter` is an abstract base class for model-specific adapters.
+2. `Flux2KleinGGUFAdapter` implements Flux2-Klein remapping + qkv split + adaLN swap.
+3. `ZImageGGUFAdapter` implements Z-Image qkv + ffn shard handling and linear qweight routing.
+4. `get_gguf_adapter(...)` strictly selects by model class/config; unsupported models raise an error (no fallback adapter).
+
+Adapter paths:
+- Base: `vllm_omni/diffusion/model_loader/gguf_adapters/base.py`
+- Z-Image: `vllm_omni/diffusion/model_loader/gguf_adapters/z_image.py`
+- Flux2-Klein: `vllm_omni/diffusion/model_loader/gguf_adapters/flux2_klein.py`
+
+## User Usage (Offline)
+
+### Baseline BF16
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model /workspace/models/black-forest-labs/FLUX.2-klein-4B \
+  --prompt "a photo of a forest with mist swirling around the tree trunks. The word 'FLUX.2' is painted over it in big, red brush strokes with visible texture" \
+  --height 768 \
+  --width 1360 \
+  --seed 42 \
+  --cfg_scale 4.0 \
+  --num_images_per_prompt 1 \
+  --num_inference_steps 4 \
+  --output outputs/flux2_klein_4b.png
+```
+
+### Native GGUF (Transformer Only)
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model /workspace/models/black-forest-labs/FLUX.2-klein-4B \
+  --gguf-model "/workspace/models/unsloth/FLUX.2-klein-4B-GGUF/flux-2-klein-4b-Q8_0.gguf" \
+  --quantization gguf \
+  --prompt "a photo of a forest with mist swirling around the tree trunks. The word 'FLUX.2' is painted over it in big, red brush strokes with visible texture" \
+  --height 768 \
+  --width 1360 \
+  --seed 42 \
+  --cfg_scale 4.0 \
+  --num_images_per_prompt 1 \
+  --num_inference_steps 4 \
+  --output outputs/flux2_klein_4b_gguf.png
+```
+
+Notes for GGUF:
+1. Many GGUF repos do not ship `model_index.json` and configs. Use the base repo for `--model` and only pass the GGUF file via `--gguf-model`.
+2. `gguf_model` supports local path, `repo/file.gguf`, or `repo:quant_type`.
+
+## User Usage (Online)
+
+### Start Server (Native GGUF via CLI)
+```bash
+vllm serve /workspace/models/black-forest-labs/FLUX.2-klein-4B \
+  --omni \
+  --port 8000 \
+  --quantization-config '{"method":"gguf","gguf_model":"/workspace/models/unsloth/FLUX.2-klein-4B-GGUF/flux-2-klein-4b-Q8_0.gguf"}'
+```
+
+### Online Request (Images API)
+```bash
+curl -X POST http://localhost:8000/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "prompt": "a dragon laying over the spine of the Green Mountains of Vermont",
+    "size": "1024x1024",
+    "seed": 42,
+    "num_inference_steps": 4
+  }'
+```

docs/user_guide/diffusion/quantization/overview.md

@@ -0,0 +1,18 @@
+# Quantization for Diffusion Transformers
+
+vLLM-Omni supports quantization of DiT linear layers to reduce memory usage and accelerate inference.
+
+## Supported Methods
+
+| Method | Guide |
+|--------|-------|
+| FP8 | [FP8](fp8.md) |
+| GGUF | [GGUF](gguf.md) |
+
+## Device Compatibility
+
+| GPU Generation | Example GPUs | FP8 Mode |
+|---------------|-------------------|----------|
+| Ada/Hopper (SM 89+) | RTX 4090, H100, H200 | Full W8A8 with native hardware |
+
+Kernel selection is automatic.

docs/user_guide/diffusion_acceleration.md

@@ -1,6 +1,6 @@
 # Diffusion Acceleration Overview
 
-vLLM-Omni supports various cache acceleration methods to speed up diffusion model inference with minimal quality degradation. These methods include **cache methods** that intelligently cache intermediate computations to avoid redundant work across diffusion timesteps, and **parallelism methods** that distribute the computation across multiple devices.
+vLLM-Omni supports various acceleration methods to speed up diffusion model inference with minimal quality degradation. These include **cache methods** that intelligently cache intermediate computations to avoid redundant work across diffusion timesteps, **parallelism methods** that distribute the computation across multiple devices, and **quantization methods** that reduce memory footprint while preserving accuracy.
 
 ## Supported Acceleration Methods
 
@@ -14,6 +14,10 @@ vLLM-Omni currently supports two main cache acceleration backends:
 
 Both methods can provide significant speedups (typically **1.5x-2.0x**) while maintaining high output quality.
 
+vLLM-Omni also supports quantization methods:
+
+3. **[FP8 Quantization](diffusion/quantization/overview.md)** - Reduces DiT linear layers from BF16 to FP8, providing ~1.28x speedup with minimal quality loss. Supports per-layer skip for sensitive layers.
+
 vLLM-Omni also supports parallelism methods for diffusion models, including:
 
 1. [Ulysses-SP](diffusion/parallelism_acceleration.md#ulysses-sp) - splits the input along the sequence dimension and uses all-to-all communication to allow each device to compute only a subset of attention heads.
@@ -22,6 +26,12 @@ vLLM-Omni also supports parallelism methods for diffusion models, including:
 
 3. [CFG-Parallel](diffusion/parallelism_acceleration.md#cfg-parallel) - runs the positive/negative prompts of classifier-free guidance (CFG) on different devices, then merges on a single device to perform the scheduler step.
 
+4. [Tensor Parallelism](diffusion/parallelism_acceleration.md#tensor-parallelism) - shards DiT weights across devices to reduce per-GPU memory usage.
+
+5. [VAE Patch Parallelism](diffusion/parallelism_acceleration.md#vae-patch-parallelism) - shards VAE decode/encode spatially across ranks to reduce VAE peak memory (and can speed up VAE decode).
+
+6. [HSDP](diffusion/parallelism_acceleration.md#hsdp) - Hybrid Sharded Data Parallel shards model weights across GPUs to reduce per-GPU memory usage, enabling inference of large models on limited GPU memory.
+
 ## Quick Comparison
 
 ### Cache Methods
@@ -31,32 +41,47 @@ vLLM-Omni also supports parallelism methods for diffusion models, including:
 | **TeaCache** | `cache_backend="tea_cache"` | Simple, adaptive caching with minimal configuration | Quick setup, balanced speed/quality |
 | **Cache-DiT** | `cache_backend="cache_dit"` | Advanced caching with multiple techniques (DBCache, TaylorSeer, SCM) | Maximum acceleration, fine-grained control |
 
+### Quantization Methods
+
+| Method | Configuration | Description | Best For |
+|--------|--------------|-------------|----------|
+| **FP8** | `quantization="fp8"` | FP8 W8A8 on Ada/Hopper, weight-only on older GPUs | Memory reduction, inference speedup |
+
 ## Supported Models
 
 The following table shows which models are currently supported by each acceleration method:
 
 ### ImageGen
 
-| Model | Model Identifier | TeaCache | Cache-DiT | Ulysses-SP | Ring-Attention | CFG-Parallel |
-|-------|------------------|:----------:|:-----------:|:-----------:|:----------------:|:----------------:|
-| **LongCat-Image** | `meituan-longcat/LongCat-Image` | ❌ | ✅ | ❌ | ❌ | ✅ |
-| **LongCat-Image-Edit** | `meituan-longcat/LongCat-Image-Edit` | ❌ | ✅ | ❌ | ❌ | ✅ |
-| **Ovis-Image** | `OvisAI/Ovis-Image` | ❌ | ✅ | ❌ | ❌ | ✅ |
-| **Qwen-Image** | `Qwen/Qwen-Image` | ✅ | ✅ | ✅ | ✅ | ✅ |
-| **Qwen-Image-2512** | `Qwen/Qwen-Image-2512` | ✅ | ✅ | ✅ | ✅ | ✅ |
-| **Qwen-Image-Edit** | `Qwen/Qwen-Image-Edit` | ✅ | ✅ | ✅ | ✅ | ✅ |
-| **Qwen-Image-Edit-2509** | `Qwen/Qwen-Image-Edit-2509` | ✅ | ✅ | ✅ | ✅ | ✅ |
-| **Qwen-Image-Layered** | `Qwen/Qwen-Image-Layered` | ❌ | ✅ | ✅ | ✅ | ✅ |
-| **Z-Image** | `Tongyi-MAI/Z-Image-Turbo` | ✅ | ✅ | ❌ | ❌ | ❌ |
-| **Stable-Diffusion3.5** | `stabilityai/stable-diffusion-3.5` | ❌ | ✅ | ❌ | ❌ | ✅ |
-| **Bagel** | `ByteDance-Seed/BAGEL-7B-MoT` | ✅ | ✅ | ❌ | ❌ | ❌ |
-| **FLUX.1-dev** | `black-forest-labs/FLUX.1-dev` | ❌ | ✅ | ❌ | ❌ | ❌ |
+| Model | Model Identifier | TeaCache | Cache-DiT | Ulysses-SP | Ring-Attention | CFG-Parallel | Tensor-Parallel | VAE-Patch-Parallel |
+|-------|------------------|:--------:|:---------:|:----------:|:--------------:|:------------:|:---------------:|:------------------:|
+| **LongCat-Image** | `meituan-longcat/LongCat-Image` | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+| **LongCat-Image-Edit** | `meituan-longcat/LongCat-Image-Edit` | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+| **Ovis-Image** | `OvisAI/Ovis-Image` | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+| **Qwen-Image** | `Qwen/Qwen-Image` | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+| **Qwen-Image-2512** | `Qwen/Qwen-Image-2512` | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+| **Qwen-Image-Edit** | `Qwen/Qwen-Image-Edit` | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+| **Qwen-Image-Edit-2509** | `Qwen/Qwen-Image-Edit-2509` | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+| **Qwen-Image-Layered** | `Qwen/Qwen-Image-Layered` | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+| **Z-Image** | `Tongyi-MAI/Z-Image-Turbo` | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ (TP=2 only) | ✅ |
+| **Stable-Diffusion3.5** | `stabilityai/stable-diffusion-3.5` | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+| **Bagel** | `ByteDance-Seed/BAGEL-7B-MoT` | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+| **FLUX.1-dev** | `black-forest-labs/FLUX.1-dev` | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+| **NextStep-1.1** | `stepfun-ai/NextStep-1.1` | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ |
 
 ### VideoGen
 
-| Model | Model Identifier | TeaCache | Cache-DiT | Ulysses-SP | Ring-Attention |CFG-Parallel |
-|-------|------------------|:--------:|:---------:|:----------:|:--------------:|:----------------:|
-| **Wan2.2** | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | ❌ | ✅ | ✅ | ✅ | ✅ |
+| Model | Model Identifier | TeaCache | Cache-DiT | Ulysses-SP | Ring-Attention | CFG-Parallel | HSDP |
+|-------|------------------|:--------:|:---------:|:----------:|:--------------:|:------------:|:----:|
+| **Wan2.2** | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ |
+
+### Quantization
+
+| Model | Model Identifier | FP8 |
+|-------|------------------|:---:|
+| **Qwen-Image** | `Qwen/Qwen-Image` | ✅ |
+| **Qwen-Image-2512** | `Qwen/Qwen-Image-2512` | ✅ |
+| **Z-Image** | `Tongyi-MAI/Z-Image-Turbo` | ✅ |
 
 
 ## Performance Benchmarks
@@ -201,6 +226,71 @@ outputs = omni.generate(
 )
 ```
 
+### Using Tensor Parallelism
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="Tongyi-MAI/Z-Image-Turbo",
+    parallel_config=DiffusionParallelConfig(tensor_parallel_size=2),
+)
+
+outputs = omni.generate(
+    prompt="a cat reading a book",
+    num_inference_steps=9,
+    width=512,
+    height=512,
+)
+```
+
+### Using VAE Patch Parallelism
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="Tongyi-MAI/Z-Image-Turbo",
+    parallel_config=DiffusionParallelConfig(
+        tensor_parallel_size=2,
+        vae_patch_parallel_size=2,
+    ),
+)
+
+outputs = omni.generate(
+    prompt="a cat reading a book",
+    num_inference_steps=9,
+    width=1024,
+    height=1024,
+)
+```
+
+### Using HSDP
+
+HSDP (Hybrid Sharded Data Parallel) shards model weights across GPUs to reduce per-GPU memory usage. This enables inference of large models (e.g., Wan2.2 14B) on GPUs with limited memory.
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+    parallel_config=DiffusionParallelConfig(
+        use_hsdp=True,
+        hsdp_replicate_size=1,  # Number of replica groups
+        hsdp_shard_size=8,      # Number of GPUs to shard across (or -1 for auto)
+    ),
+)
+
+outputs = omni.generate(
+    "A cat playing piano",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
 ### Using CFG-Parallel
 
 Run image-to-image:
@@ -227,11 +317,32 @@ outputs = omni.generate(
 )
 ```
 
+### Using FP8 Quantization
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="<your-model>",
+    quantization="fp8",
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
 ## Documentation
 
 For detailed information on each acceleration method:
 
 - **[TeaCache Guide](diffusion/teacache.md)** - Complete TeaCache documentation, configuration options, and best practices
 - **[Cache-DiT Acceleration Guide](diffusion/cache_dit_acceleration.md)** - Comprehensive Cache-DiT guide covering DBCache, TaylorSeer, SCM, and configuration parameters
+- **[FP8 Quantization Guide](diffusion/quantization/overview.md)** - FP8 quantization for DiT models with per-layer control
+- **[Tensor Parallelism](diffusion/parallelism_acceleration.md#tensor-parallelism)** - Guidance on how to enable TP for diffusion models.
 - **[Sequence Parallelism](diffusion/parallelism_acceleration.md#sequence-parallelism)** - Guidance on how to set sequence parallelism with configuration.
 - **[CFG-Parallel](diffusion/parallelism_acceleration.md#cfg-parallel)** - Guidance on how to set CFG-Parallel to run positive/negative branches across ranks.
+- **[VAE Patch Parallelism](diffusion/parallelism_acceleration.md#vae-patch-parallelism)** - Guidance on how to reduce VAE memory via patch/tile parallelism.
+- **[HSDP](diffusion/parallelism_acceleration.md#hsdp)** - Hybrid Sharded Data Parallel for memory-efficient inference of large models.

docs/user_guide/examples/offline_inference/bagel.md

@@ -154,6 +154,24 @@ The default yaml configuration deploys Thinker and DiT on the same GPU. You can
 
 ------
 
+#### Tensor Parallelism (TP)
+
+For larger models or multi-GPU environments, you can enable Tensor Parallelism (TP) by modifying the stage configuration (e.g., [`bagel.yaml`](https://github.com/vllm-project/vllm-omni/tree/main/vllm_omni/model_executor/stage_configs/bagel.yaml)).
+
+1. **Set `tensor_parallel_size`**: Increase this value (e.g., to `2` or `4`).
+2. **Set `devices`**: Specify the comma-separated GPU IDs to be used for the stage (e.g., `"0,1"`).
+
+Example configuration for TP=2 on GPUs 0 and 1:
+```yaml
+    engine_args:
+      tensor_parallel_size: 2
+      ...
+    runtime:
+      devices: "0,1"
+```
+
+------
+
 #### 🔗 Runtime Configuration
 
 | Parameter             | Value   | Description                      |
@@ -162,6 +180,74 @@ The default yaml configuration deploys Thinker and DiT on the same GPU. You can
 | `max_inflight`        | `1`     | Maximum inflight requests        |
 | `shm_threshold_bytes` | `65536` | Shared memory threshold (64KB)   |
 
+## Using Mooncake Connector
+
+[Mooncake](https://github.com/kvcache-ai/Mooncake) is a high-performance distributed KV cache transfer engine that enables efficient cross-node data movement via TCP or RDMA, making it ideal for multi-node disaggregated inference.
+
+By default, BAGEL uses `SharedMemoryConnector` for inter-stage communication. You can switch to the Mooncake connector for better performance on multi-GPU setups and to enable multi-node deployment.
+
+### Prerequisites
+
+Install the Mooncake transfer engine:
+
+```bash
+# For CUDA-enabled systems (recommended)
+pip install mooncake-transfer-engine
+
+# For non-CUDA systems
+pip install mooncake-transfer-engine-non-cuda
+```
+
+### Step 1: Start the Mooncake Master
+
+On the **primary node**, start the Mooncake master service (run in a separate terminal or background with `&`):
+
+```bash
+# Optional: enable disk-backed storage by creating a directory and passing --root_fs_dir.
+# Without it, Mooncake runs in memory-only mode, which is sufficient for KV cache transfer.
+mkdir -p ./mc_storage
+
+mooncake_master \
+  --rpc_port=50051 \
+  --enable_http_metadata_server=true \
+  --http_metadata_server_host=0.0.0.0 \
+  --http_metadata_server_port=8080 \
+  --metrics_port=9003 \
+  --root_fs_dir=./mc_storage/ \
+  --cluster_id=mc-local-1 &
+```
+
+### Step 2: Run Offline Inference with Mooncake
+
+Use the provided Mooncake stage config [`bagel_multiconnector.yaml`](https://github.com/vllm-project/vllm-omni/tree/main/vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml). Before launching, update the `metadata_server` and `master` addresses in the YAML to match your Mooncake master node's IP (use `127.0.0.1` for single-node testing).
+
+```bash
+cd examples/offline_inference/bagel
+
+# Text to Image with Mooncake
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality text2img \
+                  --prompts "A cute cat" \
+                  --stage-configs-path ../../../vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml
+
+# Image to Text with Mooncake
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality img2text \
+                  --image-path /path/to/image.jpg \
+                  --prompts "Describe this image" \
+                  --stage-configs-path ../../../vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml
+
+# Text to Text with Mooncake
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality text2text \
+                  --prompts "What is the capital of France?" \
+                  --stage-configs-path ../../../vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml
+```
+
+For more details on the Mooncake connector and multi-node setup, see the [Mooncake Store Connector documentation](https://github.com/vllm-project/vllm-omni/tree/main/docs/design/feature/omni_connectors/mooncake_store_connector.md).
+
+------
+
 ## FAQ
 
 - If you encounter an error about the backend of librosa, try to install ffmpeg with the command below.

docs/user_guide/examples/offline_inference/glm_image.md

@@ -0,0 +1,156 @@
+# GLM-Image Multistage End-to-End Inference
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/glm_image>.
+
+
+This example demonstrates how to run GLM-Image with the vLLM-Omni multistage architecture.
+
+## Architecture
+
+GLM-Image uses a 2-stage pipeline:
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                     GLM-Image Pipeline                       │
+├─────────────────────────────────────────────────────────────┤
+│                                                             │
+│  Stage 0 (AR Model)              Stage 1 (Diffusion)        │
+│  ┌─────────────────┐            ┌─────────────────────┐     │
+│  │ vLLM-optimized  │            │  GlmImagePipeline   │     │
+│  │ GlmImageFor     │  prior     │  ┌───────────────┐  │     │
+│  │ Conditional     │──tokens───►│  │ DiT Denoiser  │  │     │
+│  │ Generation      │            │  └───────────────┘  │     │
+│  │ (9B AR model)   │            │         │          │     │
+│  └─────────────────┘            │         ▼          │     │
+│         ▲                       │  ┌───────────────┐  │     │
+│         │                       │  │  VAE Decode   │──┼──► Image
+│    Text/Image                   │  └───────────────┘  │     │
+│      Input                      └─────────────────────┘     │
+│                                                             │
+└─────────────────────────────────────────────────────────────┘
+```
+
+## Features
+
+- **vLLM-optimized AR**: Uses PagedAttention and tensor parallelism for faster prior token generation
+- **Flexible deployment**: AR and Diffusion stages can run on different GPUs
+- **Text-to-Image**: Generate images from text descriptions
+- **Image-to-Image**: Edit existing images with text prompts
+
+## Usage
+
+### Text-to-Image
+
+```bash
+python end2end.py \
+    --model-path /path/to/glm-image \
+    --config-path ../../vllm_omni/model_executor/stage_configs/glm_image.yaml \
+    --prompt "A beautiful sunset over the ocean with sailing boats" \
+    --height 1024 \
+    --width 1024 \
+    --output output_t2i.png
+```
+
+### Image-to-Image (Image Editing)
+
+```bash
+python end2end.py \
+    --model-path /path/to/glm-image \
+    --config-path ../../vllm_omni/model_executor/stage_configs/glm_image.yaml \
+    --prompt "Transform this scene into a winter wonderland" \
+    --image input.png \
+    --output output_i2i.png
+```
+
+### With Custom Parameters
+
+```bash
+python end2end.py \
+    --model-path /path/to/glm-image \
+    --config-path ../../vllm_omni/model_executor/stage_configs/glm_image.yaml \
+    --prompt "A photorealistic cat sitting on a window sill" \
+    --height 1024 \
+    --width 1024 \
+    --num-inference-steps 50 \
+    --guidance-scale 1.5 \
+    --seed 42 \
+    --output output.png
+```
+
+## Shell Scripts
+
+### Run Text-to-Image
+
+```bash
+./run_t2i.sh
+```
+
+### Run Image-to-Image
+
+```bash
+./run_i2i.sh --image /path/to/input.png
+```
+
+## Stage Configuration
+
+The stage config (`glm_image.yaml`) defines:
+
+- **Stage 0 (AR)**: Uses `GPUARWorker` with vLLM engine
+
+  - Model: `GlmImageForConditionalGeneration`
+  - Output: `token_ids` (prior tokens)
+
+- **Stage 1 (Diffusion)**: Uses diffusion engine
+  - Model: `GlmImagePipeline`
+  - Output: Generated image
+
+See `vllm_omni/model_executor/stage_configs/glm_image.yaml` for full configuration.
+
+## Comparison with Single-Stage
+
+| Aspect      | Single-Stage (transformers) | Multistage (vLLM)   |
+| ----------- | --------------------------- | ------------------- |
+| AR Model    | transformers native         | vLLM PagedAttention |
+| Memory      | Higher (no KV cache opt)    | Lower (optimized)   |
+| Throughput  | Lower                       | Higher              |
+| Flexibility | Single GPU                  | Multi-GPU support   |
+
+## Troubleshooting
+
+### OOM Error
+
+Try reducing memory usage:
+
+```bash
+# In glm_image.yaml, adjust:
+gpu_memory_utilization: 0.5  # Reduce from 0.6
+```
+
+### Slow Initialization
+
+The first run loads model weights. Subsequent runs are faster:
+
+```bash
+--stage-init-timeout 900  # Increase timeout for slow storage
+```
+
+## Requirements
+
+- vLLM-Omni with GLM-Image support
+- CUDA-capable GPU (recommended: H100/A100 with 80GB)
+- GLM-Image model weights
+
+## Example materials
+
+??? abstract "end2end.py"
+    ``````py
+    --8<-- "examples/offline_inference/glm_image/end2end.py"
+    ``````
+??? abstract "run_i2i.sh"
+    ``````sh
+    --8<-- "examples/offline_inference/glm_image/run_i2i.sh"
+    ``````
+??? abstract "run_t2i.sh"
+    ``````sh
+    --8<-- "examples/offline_inference/glm_image/run_t2i.sh"
+    ``````

docs/user_guide/examples/offline_inference/image_to_image.md

@@ -22,8 +22,8 @@ python image_edit.py \
   --image qwen-bear.png \
   --prompt "Let this mascot dance under the moon, surrounded by floating stars and poetic bubbles such as 'Be Kind'" \
   --output output_image_edit.png \
-  --num_inference_steps 50 \
-  --cfg_scale 4.0
+  --num-inference-steps 50 \
+  --cfg-scale 4.0
 ```
 
 ### Multiple Image Editing (Qwen-Image-Edit-2509)
@@ -36,26 +36,26 @@ python image_edit.py \
   --image img1.png img2.png \
   --prompt "Combine these images into a single scene" \
   --output output_image_edit.png \
-  --num_inference_steps 50 \
-  --cfg_scale 4.0 \
-  --guidance_scale 1.0
+  --num-inference-steps 50 \
+  --cfg-scale 4.0 \
+  --guidance-scale 1.0
 ```
 
 Key arguments:
 
 - `--model`: model name or path. Use `Qwen/Qwen-Image-Edit-2509` or later for multiple image support.
 - `--image`: path(s) to the source image(s) (PNG/JPG, converted to RGB). Can specify multiple images.
-- `--prompt` / `--negative_prompt`: text description (string).
-- `--cfg_scale`: true classifier-free guidance scale (default: 4.0). Classifier-free guidance is enabled by setting cfg_scale > 1 and providing a negative_prompt. Higher guidance scale encourages images closely linked to the text prompt, usually at the expense of lower image quality.
-- `--guidance_scale`: guidance scale for guidance-distilled models (default: 1.0, disabled). Unlike classifier-free guidance (--cfg_scale), guidance-distilled models take the guidance scale directly as an input parameter. Enabled when guidance_scale > 1. Ignored when not using guidance-distilled models.
-- `--num_inference_steps`: diffusion sampling steps (more steps = higher quality, slower).
+- `--prompt` / `--negative-prompt`: text description (string).
+- `--cfg-scale`: true classifier-free guidance scale (default: 4.0). Classifier-free guidance is enabled by setting cfg_scale > 1 and providing a negative_prompt. Higher guidance scale encourages images closely linked to the text prompt, usually at the expense of lower image quality.
+- `--guidance-scale`: guidance scale for guidance-distilled models (default: 1.0, disabled). Unlike classifier-free guidance (--cfg-scale), guidance-distilled models take the guidance scale directly as an input parameter. Enabled when guidance_scale > 1. Ignored when not using guidance-distilled models.
+- `--num-inference-steps`: diffusion sampling steps (more steps = higher quality, slower).
 - `--output`: path to save the generated PNG.
-- `--vae_use_slicing`: enable VAE slicing for memory optimization.
-- `--vae_use_tiling`: enable VAE tiling for memory optimization.
-- `--cfg_parallel_size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
+- `--vae-use-slicing`: enable VAE slicing for memory optimization.
+- `--vae-use-tiling`: enable VAE tiling for memory optimization.
+- `--cfg-parallel-size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
 - `--enable-cpu-offload`: enable CPU offloading for diffusion models.
 
-> ℹ️ If you encounter OOM errors, try using `--vae_use_slicing` and `--vae_use_tiling` to reduce memory usage.
+> ℹ️ If you encounter OOM errors, try using `--vae-use-slicing` and `--vae-use-tiling` to reduce memory usage.
 
 ## Example materials
 

docs/user_guide/examples/offline_inference/image_to_video.md

@@ -20,15 +20,15 @@ python image_to_video.py \
   --model Wan-AI/Wan2.2-I2V-A14B-Diffusers \
   --image cherry_blossom.jpg \
   --prompt "Cherry blossoms swaying gently in the breeze, petals falling, smooth motion" \
-  --negative_prompt "<optional quality filter>" \
+  --negative-prompt "<optional quality filter>" \
   --height 480 \
   --width 832 \
-  --num_frames 48 \
-  --guidance_scale 5.0 \
-  --guidance_scale_high 6.0 \
-  --num_inference_steps 40 \
-  --boundary_ratio 0.875 \
-  --flow_shift 12.0 \
+  --num-frames 48 \
+  --guidance-scale 5.0 \
+  --guidance-scale-high 6.0 \
+  --num-inference-steps 40 \
+  --boundary-ratio 0.875 \
+  --flow-shift 12.0 \
   --fps 16 \
   --output i2v_output.mp4
 ```
@@ -40,13 +40,13 @@ python image_to_video.py \
   --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
   --image cherry_blossom.jpg \
   --prompt "Cherry blossoms swaying gently in the breeze, petals falling, smooth motion" \
-  --negative_prompt "<optional quality filter>" \
+  --negative-prompt "<optional quality filter>" \
   --height 480 \
   --width 832 \
-  --num_frames 48 \
-  --guidance_scale 4.0 \
-  --num_inference_steps 40 \
-  --flow_shift 12.0 \
+  --num-frames 48 \
+  --guidance-scale 4.0 \
+  --num-inference-steps 40 \
+  --flow-shift 12.0 \
   --fps 16 \
   --output i2v_output.mp4
 ```
@@ -57,20 +57,25 @@ Key arguments:
 - `--image`: Path to input image (required).
 - `--prompt`: Text description of desired motion/animation.
 - `--height/--width`: Output resolution (auto-calculated from image if not set). Dimensions should be multiples of 16.
-- `--num_frames`: Number of frames (default 81).
-- `--guidance_scale` and `--guidance_scale_high`: CFG scale (applied to low/high-noise stages for MoE).
-- `--negative_prompt`: Optional list of artifacts to suppress.
-- `--boundary_ratio`: Boundary split ratio for two-stage MoE models.
-- `--flow_shift`: Scheduler flow shift (5.0 for 720p, 12.0 for 480p).
-- `--num_inference_steps`: Number of denoising steps (default 50).
+- `--num-frames`: Number of frames (default 81).
+- `--guidance-scale` and `--guidance-scale-high`: CFG scale (applied to low/high-noise stages for MoE).
+- `--negative-prompt`: Optional list of artifacts to suppress.
+- `--boundary-ratio`: Boundary split ratio for two-stage MoE models.
+- `--flow-shift`: Scheduler flow shift (5.0 for 720p, 12.0 for 480p).
+- `--num-inference-steps`: Number of denoising steps (default 50).
 - `--fps`: Frames per second for the saved MP4 (requires `diffusers` export_to_video).
 - `--output`: Path to save the generated video.
-- `--vae_use_slicing`: Enable VAE slicing for memory optimization.
-- `--vae_use_tiling`: Enable VAE tiling for memory optimization.
-- `--cfg_parallel_size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
+- `--vae-use-slicing`: Enable VAE slicing for memory optimization.
+- `--vae-use-tiling`: Enable VAE tiling for memory optimization.
+- `--cfg-parallel-size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
 - `--enable-cpu-offload`: enable CPU offloading for diffusion models.
+- `--use-hsdp`: Enable Hybrid Sharded Data Parallel to shard model weights across GPUs.
+- `--hsdp-shard-size`: Number of GPUs to shard model weights across within each replica group. -1 (default) auto-calculates as world_size / replicate_size.
+- `--hsdp-replicate-size`: Number of replica groups for HSDP. Each replica holds a full sharded copy. Default 1 means pure sharding (no replication).
 
-> ℹ️ If you encounter OOM errors, try using `--vae_use_slicing` and `--vae_use_tiling` to reduce memory usage.
+
+
+> ℹ️ If you encounter OOM errors, try using `--vae-use-slicing` and `--vae-use-tiling` to reduce memory usage.
 
 ## Example materials
 

docs/user_guide/examples/offline_inference/lora_inference.md

@@ -1,108 +0,0 @@
-# LoRA Inference(Diffusion)
-
-Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/lora_inference>.
-
-
-This directory contains examples for using LoRA (Low-Rank Adaptation) adapters with vLLM-omni diffusion models for offline inference.
-The example uses the  `stabilityai/stable-diffusion-3.5-medium` as the default model, but you can replace it with other models in vLLM-omni.
-
-## Overview
-
-Similar to vLLM, vLLM-omni uses a unified LoRA handling mechanism:
-
-- **Pre-loaded LoRA**: Loaded at initialization via `--lora-path` (pre-loaded into cache)
-- **Per-request LoRA**: Loaded on-demand. In the example, the LoRA is loaded via `--lora-request-path` in each request
-
-Both approaches use the same underlying mechanism - all LoRA adapters are handled uniformly through `set_active_adapter()`. If no LoRA request is provided in a request, all adapters are deactivated.
-
-## Usage
-
-### Pre-loaded LoRA (via --lora-path)
-
-Load a LoRA adapter at initialization. This adapter is pre-loaded into the cache and can be activated by requests:
-
-```bash
-python -m examples.offline_inference.lora_inference.lora_inference \
-    --prompt "A piece of cheesecake" \
-    --lora-path /path/to/lora/ \
-    --lora-scale 1.0 \
-    --num_inference_steps 50 \
-    --height 1024 \
-    --width 1024 \
-    --output output_preloaded.png
-```
-
-**Note**: When using `--lora-path`, the adapter is loaded at init time with a stable ID derived from the adapter path. This example activates it automatically for the request.
-
-### Per-request LoRA (via --lora-request-path)
-
-Load a LoRA adapter on-demand for each request:
-
-```bash
-python -m examples.offline_inference.lora_inference.lora_inference \
-    --prompt "A piece of cheesecake" \
-    --lora-request-path /path/to/lora/ \
-    --lora-scale 1.0 \
-    --num_inference_steps 50 \
-    --height 1024 \
-    --width 1024 \
-    --output output_per_request.png
-```
-
-### No LoRA
-
-If no LoRA request is provided, we will use the base model without any LoRA adapters:
-
-```bash
-python -m examples.offline_inference.lora_inference.lora_inference \
-    --prompt "A piece of cheesecake" \
-    --num_inference_steps 50 \
-    --height 1024 \
-    --width 1024 \
-    --output output_no_lora.png
-```
-
-## Parameters
-
-### LoRA Parameters
-
-- `--lora-path`: Path to LoRA adapter folder to pre-load at initialization (loads into cache with a stable ID derived from the path)
-- `--lora-request-path`: Path to LoRA adapter folder for per-request loading
-- `--lora-request-id`: Integer ID for the LoRA adapter (optional). If not provided and `--lora-request-path` is set, will derive a stable ID from the path.
-- `--lora-scale`: Scale factor for LoRA weights (default: 1.0). Higher values increase the influence of the LoRA adapter.
-
-### Standard Parameters
-
-- `--prompt`: Text prompt for image generation (required)
-- `--seed`: Random seed for reproducibility (default: 42)
-- `--height`: Image height in pixels (default: 1024)
-- `--width`: Image width in pixels (default: 1024)
-- `--num_inference_steps`: Number of denoising steps (default: 50)
-- `--output`: Output file path (default: `lora_output.png`)
-
-## How LoRA Works
-
-All LoRA adapters are handled uniformly:
-
-1. **Initialization**: If `--lora-path` is provided, the adapter is loaded into cache with a stable ID derived from the adapter path
-2. **Per-request**: If `--lora-request-path` is provided, the adapter is loaded/activated for that request
-3. **No LoRA**: If no LoRA request is provided (`req.lora_request` is None), all adapters are deactivated
-
-The system uses LRU cache management - adapters are cached and evicted when the cache is full (unless pinned).
-
-## LoRA Adapter Format
-
-LoRA adapters must be in PEFT (Parameter-Efficient Fine-Tuning) format. A typical LoRA adapter directory structure:
-
-```
-lora_adapter/
-├── adapter_config.json
-└── adapter_model.safetensors
-```
-
-## Example materials
-
-??? abstract "lora_inference.py"
-    ``````py
-    --8<-- "examples/offline_inference/lora_inference/lora_inference.py"
-    ``````

docs/user_guide/examples/offline_inference/qwen3_tts.md

@@ -90,13 +90,43 @@ Examples:
 python end2end.py --query-type Base --mode-tag icl
 ```
 
+## Streaming Mode
+
+Add `--streaming` to stream audio chunks progressively via `AsyncOmni` (requires `async_chunk: true` in the stage config):
+
+```bash
+python end2end.py --query-type CustomVoice --streaming --output-dir /tmp/out_stream
+```
+
+Each 25-frame Code2Wav chunk is logged as it arrives. The final WAV file is written once generation
+completes. This demonstrates that audio data is available progressively rather than only at the end.
+
+> **Note:** Streaming uses `AsyncOmni` internally. The non-streaming path (`Omni`) is unchanged.
+
+## Batched Decoding
+
+The Code2Wav stage (stage 1) supports batched decoding, where multiple requests are decoded in a single forward pass through the SpeechTokenizer. To use it, provide a stage config with `max_batch_size > 1` and pass multiple prompts via `--txt-prompts` with a matching `--batch-size`.
+
+```
+python end2end.py --query-type CustomVoice \
+    --txt-prompts benchmark_prompts.txt \
+    --batch-size 4 \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts_batch.yaml
+```
+
+**Important:** `--batch-size` must match a CUDA graph capture size (1, 2, 4, 8, 16...) because the Talker's code predictor KV cache is sized to `max_num_seqs`, and CUDA graphs pad the batch to the next capture size. Both stages need `max_batch_size >= batch_size` in the stage config for batching to take effect. If only stage 1 has a higher `max_batch_size`, it won't help — stage 1 can only batch chunks from requests that are in-flight simultaneously, which requires stage 0 to also process multiple requests concurrently.
+
 ## Notes
 
 - The script uses the model paths embedded in `end2end.py`. Update them if your local cache path differs.
-- Use `--output-dir` (preferred) or `--output-wav` to change the output folder.
+- Use `--output-dir` to change the output folder.
 
 ## Example materials
 
+??? abstract "benchmark_prompts.txt"
+    ``````txt
+    --8<-- "examples/offline_inference/qwen3_tts/benchmark_prompts.txt"
+    ``````
 ??? abstract "end2end.py"
     ``````py
     --8<-- "examples/offline_inference/qwen3_tts/end2end.py"

docs/user_guide/examples/offline_inference/text_to_audio.md

@@ -21,22 +21,22 @@ If you use a gated model (e.g., `stabilityai/stable-audio-open-1.0`), ensure you
 python text_to_audio.py \
   --model stabilityai/stable-audio-open-1.0 \
   --prompt "The sound of a hammer hitting a wooden surface" \
-  --negative_prompt "Low quality" \
+  --negative-prompt "Low quality" \
   --seed 42 \
-  --guidance_scale 7.0 \
-  --audio_length 10.0 \
-  --num_inference_steps 100 \
+  --guidance-scale 7.0 \
+  --audio-length 10.0 \
+  --num-inference-steps 100 \
   --output stable_audio_output.wav
 ```
 
 Key arguments:
 
 - `--prompt`: text description (string).
-- `--negative_prompt`: negative prompt for classifier-free guidance.
+- `--negative-prompt`: negative prompt for classifier-free guidance.
 - `--seed`: integer seed for deterministic generation.
-- `--guidance_scale`: classifier-free guidance scale.
-- `--audio_length`: audio duration in seconds.
-- `--num_inference_steps`: diffusion sampling steps.(more steps = higher quality, slower).
+- `--guidance-scale`: classifier-free guidance scale.
+- `--audio-length`: audio duration in seconds.
+- `--num-inference-steps`: diffusion sampling steps.(more steps = higher quality, slower).
 - `--output`: path to save the generated WAV file.
 
 ## Example materials

docs/user_guide/examples/offline_inference/text_to_image.md

@@ -3,7 +3,7 @@
 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_image>.
 
 
-This folder provides several entrypoints for experimenting with `Qwen/Qwen-Image` `Qwen/Qwen-Image-2512` `Tongyi-MAI/Z-Image-Turbo` using vLLM-Omni:
+This folder provides several entrypoints for experimenting with `Qwen/Qwen-Image` `Qwen/Qwen-Image-2512` `Tongyi-MAI/Z-Image-Turbo` `stepfun-ai/NextStep-1.1` using vLLM-Omni, note that NextStep-1.1 has different architecture so we treat it differently regarding running arguments and pipeline.
 
 - `text_to_image.py`: command-line script for single image generation with advanced options.
 - `web_demo.py`: lightweight Gradio UI for interactive prompt/seed/CFG exploration.
@@ -51,8 +51,7 @@ if __name__ == "__main__":
 
     For diffusion pipelines, the stage config field `stage_args.[].runtime.max_batch_size` is 1 by default, and the input
     list is sliced into single-item requests before feeding into the diffusion pipeline. For models that do internally support
-    batched inputs, you can [modify this configuration](../../../configuration/stage_configs.md
-    ) to let the model accept a longer batch of prompts.
+    batched inputs, you can [modify this configuration](https://github.com/vllm-project/vllm-omni/tree/main/configuration/stage_configs.md) to let the model accept a longer batch of prompts.
 
 Apart from string prompt, vLLM-Omni also supports dictionary prompts in the same style as vLLM.
 This is useful for models that support negative prompts.
@@ -78,37 +77,95 @@ if __name__ == "__main__":
 
 ## Local CLI Usage
 
+### Qwen/Tongyi Models
+
 ```bash
 python text_to_image.py \
   --model Tongyi-MAI/Z-Image-Turbo \
   --prompt "a cup of coffee on the table" \
   --seed 42 \
-  --cfg_scale 4.0 \
-  --num_images_per_prompt 1 \
-  --num_inference_steps 50 \
+  --cfg-scale 4.0 \
+  --num-images-per-prompt 1 \
+  --num-inference-steps 50 \
   --height 1024 \
   --width 1024 \
   --output outputs/coffee.png
 ```
 
-Key arguments:
+### NextStep Models
+
+NextStep-1.1 can have extra arguments
+```bash
+python text_to_image.py \
+  --model stepfun-ai/NextStep-1.1 \
+  --prompt "A baby panda wearing an Iron Man mask, holding a board with 'NextStep-1' written on it" \
+  --height 512 \
+  --width 512 \
+  --num-inference-steps 28 \
+  --guidance-scale 7.5 \
+  --guidance-scale-2 1.0 \
+  --cfg-schedule constant \
+  --output nextstep_output.png \
+  --seed 42
+```
+
+### Key Arguments
+
+**Common arguments:**
 
 - `--prompt`: text description (string).
 - `--seed`: integer seed for deterministic sampling.
-- `--cfg_scale`: true CFG scale (model-specific guidance strength).
-- `--num_images_per_prompt`: number of images to generate per prompt (saves as `output`, `output_1`, ...).
-- `--num_inference_steps`: diffusion sampling steps (more steps = higher quality, slower).
+- `--cfg-scale`: true CFG scale (model-specific guidance strength).
+- `--num-images-per-prompt`: number of images to generate per prompt (saves as `output`, `output_1`, ...).
+- `--num-inference-steps`: diffusion sampling steps (more steps = higher quality, slower).
 - `--height/--width`: output resolution (defaults 1024x1024).
 - `--output`: path to save the generated PNG.
-- `--vae_use_slicing`: enable VAE slicing for memory optimization.
-- `--vae_use_tiling`: enable VAE tiling for memory optimization.
-- `--cfg_parallel_size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
+- `--vae-use-slicing`: enable VAE slicing for memory optimization.
+- `--vae-use-tiling`: enable VAE tiling for memory optimization.
+- `--cfg-parallel-size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion_acceleration.md#using-cfg-parallel).
 - `--enable-cpu-offload`: enable CPU offloading for diffusion models.
+- `--guidance-scale`: classifier-free guidance scale.
 
-> ℹ️ If you encounter OOM errors, try using `--vae_use_slicing` and `--vae_use_tiling` to reduce memory usage.
+**NextStep-1.1 specific:**
+- `--guidance-scale-2`: secondary guidance scale, e.g. image-level CFG (default: 1.0).
+- `--timesteps-shift`: timesteps shift parameter for sampling (default: 1.0).
+- `--cfg-schedule`: CFG schedule type, "constant" or "linear" (default: "constant").
+- `--use-norm`: apply layer normalization to sampled tokens.
+
+> ℹ️ If you encounter OOM errors, try using `--vae-use-slicing` and `--vae-use-tiling` to reduce memory usage.
 
 > ℹ️ Qwen-Image currently publishes best-effort presets at `1328x1328`, `1664x928`, `928x1664`, `1472x1140`, `1140x1472`, `1584x1056`, and `1056x1584`. Adjust `--height/--width` accordingly for the most reliable outcomes.
 
+## LoRA
+
+This example supports Peft-compatible LoRA (Low-Rank Adaptation) adapters for diffusion models. Pass `--lora-path` to use a LoRA adapter and optionally `--lora-scale` (default 1.0); omit it to use the base model only.
+
+### Basic usage with LoRA
+
+```bash
+python text_to_image.py \
+  --model Tongyi-MAI/Z-Image-Turbo \
+  --prompt "A piece of cheesecake" \
+  --lora-path /path/to/lora/ \
+  --lora-scale 1.0 \
+  --output output.png
+```
+
+### LoRA parameters
+
+- `--lora-path`: Path to LoRA adapter folder (PEFT format). Loaded at initialization and used for generation.
+- `--lora-scale`: Scale factor for LoRA weights (default: 1.0). Higher values increase the influence of the LoRA adapter.
+
+### LoRA adapter format
+
+LoRA adapters must be in PEFT (Parameter-Efficient Fine-Tuning) format. A typical LoRA adapter directory structure:
+
+```
+lora_adapter/
+├── adapter_config.json
+└── adapter_model.safetensors
+```
+
 ## Web UI Demo
 
 Launch the gradio demo:

docs/user_guide/examples/offline_inference/text_to_video.md

@@ -10,14 +10,14 @@ The `Wan-AI/Wan2.2-T2V-A14B-Diffusers` pipeline generates short videos from text
 ```bash
 python text_to_video.py \
   --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
-  --negative_prompt "<optional quality filter>" \
+  --negative-prompt "<optional quality filter>" \
   --height 480 \
   --width 832 \
-  --num_frames 33 \
-  --guidance_scale 4.0 \
-  --guidance_scale_high 3.0 \
-  --flow_shift 12.0 \
-  --num_inference_steps 40 \
+  --num-frames 33 \
+  --guidance-scale 4.0 \
+  --guidance-scale-high 3.0 \
+  --flow-shift 12.0 \
+  --num-inference-steps 40 \
   --fps 16 \
   --output t2v_out.mp4
 ```
@@ -26,18 +26,18 @@ Key arguments:
 
 - `--prompt`: text description (string).
 - `--height/--width`: output resolution (defaults 480x832, i.e. 480P). Dimensions should align with Wan VAE downsampling (multiples of 8).
-- `--num_frames`: Number of frames (Wan default is 81).
-- `--guidance_scale` and `--guidance_scale_high`: CFG scale (applied to low/high).
-- `--negative_prompt`: optional list of artifacts to suppress (the PR demo used a long Chinese string).
-- `--boundary_ratio`: Boundary split ratio for low/high DiT. Default `0.875` uses both transformers for best quality. Set to `1.0` to load only the low-noise transformer (saves noticeable memory with good quality, recommended if memory is limited). Set to `0.0` loads only the high-noise transformer (not recommended, lower quality).
+- `--num-frames`: Number of frames (Wan default is 81).
+- `--guidance-scale` and `--guidance-scale-high`: CFG scale (applied to low/high).
+- `--negative-prompt`: optional list of artifacts to suppress (the PR demo used a long Chinese string).
+- `--boundary-ratio`: Boundary split ratio for low/high DiT. Default `0.875` uses both transformers for best quality. Set to `1.0` to load only the low-noise transformer (saves noticeable memory with good quality, recommended if memory is limited). Set to `0.0` loads only the high-noise transformer (not recommended, lower quality).
 - `--fps`: frames per second for the saved MP4 (requires `diffusers` export_to_video).
 - `--output`: path to save the generated video.
-- `--vae_use_slicing`: enable VAE slicing for memory optimization.
-- `--vae_use_tiling`: enable VAE tiling for memory optimization.
-- `--cfg_parallel_size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
+- `--vae-use-slicing`: enable VAE slicing for memory optimization.
+- `--vae-use-tiling`: enable VAE tiling for memory optimization.
+- `--cfg-parallel-size`: set it to 2 to enable CFG Parallel. See more examples in [`user_guide`](https://github.com/vllm-project/vllm-omni/tree/main/docs/user_guide/diffusion/parallelism_acceleration.md#cfg-parallel).
 - `--enable-cpu-offload`: enable CPU offloading for diffusion models.
 
-> ℹ️ If you encounter OOM errors, try using `--vae_use_slicing` and `--vae_use_tiling` to reduce memory usage.
+> ℹ️ If you encounter OOM errors, try using `--vae-use-slicing` and `--vae-use-tiling` to reduce memory usage.
 
 ## Example materials
 

docs/user_guide/examples/online_serving/bagel.md

@@ -25,12 +25,152 @@ cd /workspace/vllm-omni/examples/online_serving/bagel
 bash run_server.sh
 ```
 
-If you have a custom stage configs file, launch the server with the command below:
-
 ```bash
 vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
 ```
 
+#### 🚀 Tensor Parallelism (TP)
+
+For larger models or multi-GPU environments, you can enable Tensor Parallelism (TP) for the server.
+
+1. **Modify Stage Config**: Create or modify a stage configuration yaml (e.g., [`bagel.yaml`](https://github.com/vllm-project/vllm-omni/tree/main/vllm_omni/model_executor/stage_configs/bagel.yaml)). Set `tensor_parallel_size` to `2` (or more) and update `devices` to include multiple GPU IDs (e.g., `"0,1"`).
+
+```yaml
+    engine_args:
+      tensor_parallel_size: 2
+      ...
+    runtime:
+      devices: "0,1"
+```
+
+2. **Launch Server**:
+```bash
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 --stage-configs-path /path/to/your/custom_bagel.yaml
+```
+
+#### Using Mooncake Connector
+
+By default, BAGEL uses `SharedMemoryConnector` for inter-stage communication. You can use the [Mooncake](https://github.com/kvcache-ai/Mooncake) connector to transfer KV cache between stages, which also enables multi-node deployment.
+
+**1. Install Mooncake**
+
+```bash
+# For CUDA-enabled systems (recommended)
+pip install mooncake-transfer-engine
+
+# For non-CUDA systems
+pip install mooncake-transfer-engine-non-cuda
+```
+
+**2. Start Mooncake Master** on the primary node:
+
+```bash
+# Optional: enable disk-backed storage by creating a directory and passing --root_fs_dir.
+# Without it, Mooncake runs in memory-only mode, which is sufficient for KV cache transfer.
+mkdir -p ./mc_storage
+
+mooncake_master \
+  --rpc_port=50051 \
+  --enable_http_metadata_server=true \
+  --http_metadata_server_host=0.0.0.0 \
+  --http_metadata_server_port=8080 \
+  --metrics_port=9003 \
+  --root_fs_dir=./mc_storage/ \
+  --cluster_id=mc-local-1 &
+```
+
+**3. Launch the server** with the Mooncake stage config:
+
+```bash
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml
+```
+
+> **Note**: Before launching, edit [`bagel_multiconnector.yaml`](https://github.com/vllm-project/vllm-omni/tree/main/vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml) and replace the `metadata_server` and `master` addresses with your Mooncake master node's actual IP. For single-node testing, `127.0.0.1` works.
+
+The client-side usage is identical to the default setup -- the Mooncake connector is transparent to the API. See the requests section below.
+
+For more details on the Mooncake connector configuration, see the [Mooncake Store Connector documentation](https://github.com/vllm-project/vllm-omni/tree/main/docs/design/feature/omni_connectors/mooncake_store_connector.md).
+
+#### Multi-Node Deployment
+
+You can deploy each stage on a **separate node** for better resource utilization. In this example, the orchestrator (Stage 0 / Thinker) and Stage 1 (DiT) run on different machines, connected via Mooncake.
+
+Replace `<ORCHESTRATOR_IP>` below with the actual IP address of your orchestrator node (e.g., `10.244.227.244`).
+
+> [!WARNING]
+> **Before launching**, edit [`bagel_multiconnector.yaml`](https://github.com/vllm-project/vllm-omni/tree/main/vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml) and replace the `metadata_server` and `master` addresses with your Mooncake master node's actual IP. Mismatched addresses will cause silent connection failures.
+
+**1. Start Mooncake Master** (on the orchestrator node):
+
+```bash
+mooncake_master \
+  --rpc_port=50051 \
+  --enable_http_metadata_server=true \
+  --http_metadata_server_host=<ORCHESTRATOR_IP> \
+  --http_metadata_server_port=8080 \
+  --metrics_port=9003
+```
+
+**2. Launch Stage 0 (Thinker / Orchestrator)** on the orchestrator node:
+
+```bash
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni \
+    --port 8000 \ # API server port for client requests
+    --stage-configs-path vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml \
+    --stage-id 0 \
+    -oma <ORCHESTRATOR_IP> \
+    -omp 8091
+```
+
+**3. Launch Stage 1 (DiT)** on the remote node in headless mode:
+
+```bash
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/bagel_multiconnector.yaml \
+    --stage-id 1 \
+    --headless \
+    -oma <ORCHESTRATOR_IP> \
+    -omp 8091
+```
+
+**Mooncake Master arguments:**
+
+| Argument | Description |
+| :------- | :---------- |
+| `--rpc_port` | Mooncake RPC port for control-plane coordination between stages |
+| `--enable_http_metadata_server` | Enable the HTTP metadata server for service discovery |
+| `--http_metadata_server_host` | IP address to bind the metadata server (use the orchestrator node's IP) |
+| `--http_metadata_server_port` | Port for the HTTP metadata server |
+| `--metrics_port` | Port for Prometheus-compatible metrics endpoint |
+
+**vllm serve arguments:**
+
+| Argument | Description |
+| :------- | :---------- |
+| `--stage-id` | Which stage this process runs (0 = Thinker, 1 = DiT) |
+| `--headless` | Run without the API server (worker-only mode) |
+| `-oma` | Orchestrator master address |
+| `-omp` | Orchestrator master port for Stage 1 to connect to Stage 0 for task coordination |
+
+> [!IMPORTANT]
+> **Startup Order**: Stage 0 (orchestrator) must be launched **before** Stage 1 (headless).
+> Stage 0 will appear to hang on startup until Stage 1 (worker) connects — this is expected behavior.
+
+**Network Requirements**
+
+All nodes must have network connectivity to each other. Ensure the following ports are open **between all participating nodes**:
+
+| Port | Protocol | Service | Direction |
+| :--- | :------- | :------ | :-------- |
+| 50051 | TCP | Mooncake Master RPC | Worker → Orchestrator |
+| 8080 | TCP | Mooncake HTTP Metadata Server | Worker → Orchestrator |
+| 8091 | TCP | Orchestrator Master (`-omp`) | Worker → Orchestrator |
+| 8000 | TCP | API Server (`--port`) | Client → Orchestrator |
+| 9003 | TCP | Metrics (optional) | Monitoring → Orchestrator |
+
+> **Tip**: If nodes are behind a firewall or in different VPCs/security groups, make sure the above ports are allowed in ingress/egress rules. All nodes should be reachable via their IP addresses (no NAT). Using nodes on the same subnet or VPC is recommended to minimize latency for Mooncake KV cache transfers.
+
 ### Send Multi-modal Request
 
 Get into the bagel folder:
@@ -242,3 +382,7 @@ sudo apt install ffmpeg
     ``````sh
     --8<-- "examples/online_serving/bagel/run_server.sh"
     ``````
+??? abstract "run_server_stage_cli.sh"
+    ``````sh
+    --8<-- "examples/online_serving/bagel/run_server_stage_cli.sh"
+    ``````