ci: raise merge-tier entrypoint timeouts (H100 60->90, L4 30->45)

Both entrypoint jobs are killed mid-run by timeout_in_minutes while tests
are still passing, not by any failing test. The budgets had no headroom:
on main the H100 job timed out in 2 of the last 3 gate builds (2648, 2650)
and passed 2651 at 56/60 min; the L4 job passed at 26/30 min only with a
warm image cache. On the vLLM 0.25 wheel, model weight loading is slower
on the CI runners (like-for-like: +10 min total on H100, 1.7B TTS load
11s -> 132s on L4), which turns the coin flip into a guaranteed timeout
(build 2655).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

.buildkite/pipeline-intel.yaml

@@ -10,7 +10,7 @@ steps:
           DOCKER_BUILDKIT: "1"
           # Buildkite will automatically replace this with the actual commit hash
           VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
-          VLLM_VERSION: "v0.23.0"
+          VLLM_VERSION: "v0.24.0"
         priority: 100
         timeout_in_minutes: 60
         soft_fail: false

.buildkite/test-amd-merge.yml

@@ -7,17 +7,29 @@ steps:
 
 - group: ":card_index_dividers: Simple Test"
   steps:
-    - label: "Simple · Diffusion & Model Executor Test"
+    - label: "Simple · Diffusion & Model Executor Test (1/2)"
       agent_pool: mi325_1
       depends_on: amd-build
-      timeout_in_minutes: 50
+      timeout_in_minutes: 70
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_ROCM_USE_AITER=0
         # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
         # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        # CPU tests on ROCm gpu machines are very slow, so we give them more time.
+        - "timeout 60m pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+
+    - label: "Simple · Diffusion & Model Executor Test (2/2)"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      timeout_in_minutes: 70
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        # CPU tests on ROCm gpu machines are very slow, so we give them more time.
+        - "timeout 60m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
     - label: "Simple · Other Test"
       agent_pool: mi325_1
@@ -158,7 +170,7 @@ steps:
       grade: Blocking
       commands:
         - |
-          timeout 20m bash -c '
+          timeout 50m bash -c '
             export VLLM_LOGGING_LEVEL=DEBUG
             export VLLM_WORKER_MULTIPROC_METHOD=spawn
             export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
@@ -170,39 +182,14 @@ steps:
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
-      commands:
-        - |
-          timeout 30m bash -c '
-            export VLLM_LOGGING_LEVEL=DEBUG
-            export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model"
-          '
-
-    # It is the whisper ASR validation step that is slow, not the Voxtral TTS generation.
-    # whisper ASR runs on CPU.
-    - label: "Voxtral-TTS E2E Test (Online Serving)"
-      agent_pool: mi325_1
-      depends_on: amd-build
-      mirror_hardwares: [amdproduction]
-      grade: Blocking
+      timeout_in_minutes: 50
       commands:
         - |
           timeout 40m bash -c '
+            export VLLM_LOGGING_LEVEL=DEBUG
             export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-          '
-
-    - label: "Voxtral-TTS E2E Test (Offline Inference)"
-      agent_pool: mi325_1
-      depends_on: amd-build
-      mirror_hardwares: [amdproduction]
-      grade: Blocking
-      commands:
-        - |
-          timeout 30m bash -c '
-            export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
           '
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa

.buildkite/test-amd-ready.yaml

@@ -161,19 +161,6 @@ steps:
 
 - group: ":card_index_dividers: E2E Test"
   steps:
-    - label: "MiMo-Audio E2E Test with H100"
-      agent_pool: mi325_1
-      depends_on: amd-build
-      mirror_hardwares: [amdproduction]
-      grade: Blocking
-      commands:
-        - |
-          timeout 30m bash -c '
-            export VLLM_LOGGING_LEVEL=DEBUG
-            export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m "core_model" --run-level "core_model"
-          '
-
     - label: "Qwen3-TTS E2E Test"
       agent_pool: mi325_1
       depends_on: amd-build
@@ -193,7 +180,7 @@ steps:
       commands:
         - |
           timeout 20m bash -c '
-            pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "full_model" --run-level "core_model"
+            pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"
           '
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa

.buildkite/test-merge.yml

@@ -61,13 +61,14 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
   - label: "Distributed Test"
+    depends_on: upload-merge-pipeline
     timeout_in_minutes: 60
     commands:
       - |
         timeout 55m bash -c "
           set -e
           export VLLM_IMAGE_FETCH_TIMEOUT=60
-          pytest -s -v tests/distributed -m 'advanced_model and cuda' --run-level 'advanced_model'
+          pytest -s -v tests/distributed -m 'advanced_model and cuda and H100' --run-level 'advanced_model'
         "
     agents:
       queue: "mithril-h100-pool"
@@ -107,7 +108,10 @@ steps:
     depends_on: upload-merge-pipeline
     steps:
       - label: "Entrypoint Test with H100"
-        timeout_in_minutes: 60
+        # The suite relaunches an OmniServer per test class (~30 model loads,
+        # incl. ten ~27.5 GiB Bagel loads); 60 min passed only ~1 in 3 runs
+        # even on main, and vLLM 0.25 weight loading needs ~10 min more.
+        timeout_in_minutes: 90
         commands:
           - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and H100" --run-level "advanced_model"
         agents:
@@ -145,7 +149,9 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Entrypoint Test with L4"
-        timeout_in_minutes: 30
+        # 30 min left no headroom for a cold image pull (~11 min for new base
+        # layers) plus per-class engine relaunches; main passed at 26/30.
+        timeout_in_minutes: 45
         commands:
           - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and L4" --run-level "advanced_model"
         agents:
@@ -164,28 +170,6 @@ steps:
   - group: ":card_index_dividers: E2E Test"
     depends_on: upload-merge-pipeline
     steps:
-      - label: "TTS · Step-Audio2 E2E Test"
-        timeout_in_minutes: 20
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/offline_inference/test_step_audio2.py -m 'advanced_model' --run-level 'advanced_model'
-            "
-        agents:
-          queue: "gpu_4_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
       - label: "TTS · Qwen3-TTS CustomVoice Test"
         source_file_dependencies:
           - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
@@ -195,7 +179,7 @@ steps:
           - vllm_omni/deploy/qwen3_tts.yaml
         commands:
           - |
-            timeout 20m bash -c "
+            timeout 40m bash -c "
               export VLLM_LOGGING_LEVEL=DEBUG
               export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m 'advanced_model and cuda' --run-level 'advanced_model'
@@ -222,7 +206,7 @@ steps:
           - vllm_omni/deploy/qwen3_tts.yaml
         commands:
           - |
-            timeout 20m bash -c "
+            timeout 40m bash -c "
               export VLLM_LOGGING_LEVEL=DEBUG
               export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
@@ -240,19 +224,19 @@ steps:
               volumes:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
-      - label: "TTS · Ming-Omni-TTS Test"
+      - label: "TTS · Higgs-Audio-V3 Test"
         source_file_dependencies:
-          - tests/e2e/online_serving/test_ming_tts.py
-          - tests/e2e/offline_inference/test_ming_tts.py
-          - vllm_omni/model_executor/models/ming_tts/
-          - vllm_omni/model_executor/stage_input_processors/ming_tts.py
-          - vllm_omni/deploy/ming_tts.yaml
-        timeout_in_minutes: 20
+          - tests/e2e/online_serving/test_higgs_audio_v3.py
+          - vllm_omni/model_executor/models/higgs_audio_v3/
+          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
+          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
         commands:
           - |
             timeout 20m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              pytest -s -v tests/e2e/online_serving/test_ming_tts.py tests/e2e/offline_inference/test_ming_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              export VLLM_USE_DEEP_GEMM=0
+              export VLLM_MOE_USE_DEEP_GEMM=0
+              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'advanced_model' --run-level 'advanced_model'
             "
         agents:
           queue: "gpu_1_queue"
@@ -264,105 +248,10 @@ steps:
               shm-size: "8gb"
               environment:
                 - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
               volumes:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
-      - label: "TTS · Voxtral-TTS Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_voxtral_tts.py
-          - tests/e2e/online_serving/test_voxtral_tts.py
-          - vllm_omni/model_executor/models/voxtral_tts/
-          - vllm_omni/model_executor/stage_input_processors/voxtral_tts.py
-          - vllm_omni/deploy/voxtral_tts.yaml
-        timeout_in_minutes: 20
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py tests/e2e/online_serving/test_voxtral_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
-      - label: "TTS · GLM-TTS Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_glm_tts.py
-          - tests/e2e/offline_inference/test_glm_tts.py
-          - vllm_omni/model_executor/models/glm_tts/
-          - vllm_omni/model_executor/stage_input_processors/glm_tts.py
-          - vllm_omni/deploy/glm_tts.yaml
-        timeout_in_minutes: 20
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              pytest -s -v tests/e2e/online_serving/test_glm_tts.py tests/e2e/offline_inference/test_glm_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
       - label: "TTS · IndexTTS2 Test"
         source_file_dependencies:
           - tests/e2e/online_serving/test_indextts2.py
@@ -378,10 +267,10 @@ steps:
           - vllm_omni/entrypoints/openai/tts_adapters/indextts2.py
           - vllm_omni/config/pipeline_registry.py
           - vllm_omni/model_executor/models/registry.py
-        timeout_in_minutes: 20
+        timeout_in_minutes: 60
         commands:
           - |
-            timeout 20m bash -c "
+            timeout 45m bash -c "
               export VLLM_LOGGING_LEVEL=DEBUG
               pytest -s -v tests/e2e/online_serving/test_indextts2.py tests/e2e/offline_inference/test_indextts2.py -m 'advanced_model and cuda' --run-level 'advanced_model'
             "
@@ -406,7 +295,7 @@ steps:
           - vllm_omni/model_executor/models/qwen3_omni/
           - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
           - vllm_omni/deploy/qwen3_omni_moe.yaml
-        timeout_in_minutes: 30
+        timeout_in_minutes: 50
         commands:
           - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
         agents:
@@ -443,49 +332,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: "Omni · Mimo Audio Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_mimo_audio.py
-          - vllm_omni/model_executor/models/mimo_audio/
-          - vllm_omni/model_executor/stage_input_processors/mimo_audio.py
-          - vllm_omni/deploy/mimo_audio.yaml
-        timeout_in_minutes: 30
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 2
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
       - label: "TTS · VoxCPM2 Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_voxcpm2_tts.py
@@ -514,37 +360,6 @@ steps:
               volumes:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
-      - label: "Omni · Higgs-Audio V2 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v2.py
-          - tests/e2e/offline_inference/test_higgs_audio_v2.py
-          - vllm_omni/model_executor/models/higgs_audio_v2/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v2.py
-          - vllm_omni/deploy/higgs_audio_v2.yaml
-        timeout_in_minutes: 30
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v2.py tests/e2e/offline_inference/test_higgs_audio_v2.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
       - label: "TTS · SoulX-Singer Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_soulxsinger.py
@@ -614,88 +429,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: "Diffusion · Longcat Image Edit Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_longcat_image_edit.py
-          - vllm_omni/diffusion/models/longcat_image/
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_longcat_image_edit.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
-      - label: "Diffusion · Longcat Image Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_longcat_image.py
-          - vllm_omni/diffusion/models/longcat_image/
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_longcat_image.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
       - label: "Diffusion · Bagel Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_bagel.py
@@ -875,50 +608,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: "Diffusion · HunyuanVideo 1.5 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_hunyuan_video_15.py
-          - vllm_omni/diffusion/models/hunyuan_video/
-        timeout_in_minutes: 40
-        commands:
-          - |
-            timeout 40m bash -c "
-              pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-            "
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
       - label: "Diffusion · Z Image Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_z_image.py
@@ -943,11 +632,10 @@ steps:
       - label: "Diffusion · FLUX.2 Klein Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_flux2_klein.py
-          - tests/e2e/online_serving/test_flux2_klein.py
           - vllm_omni/diffusion/models/flux2_klein/
         timeout_in_minutes: 30
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_flux2_klein.py tests/e2e/online_serving/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+          - pytest -s -v tests/e2e/offline_inference/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
         agents:
           queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
         plugins:
@@ -960,48 +648,3 @@ steps:
                 - "HF_TOKEN"
               volumes:
                 - "/fsx/hf_cache:/fsx/hf_cache"
-
-      - label: "Diffusion · SD3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_sd3.py
-          - vllm_omni/diffusion/models/sd3/
-        timeout_in_minutes: 45
-        commands:
-          - |
-            timeout 45m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_sd3.py -m 'advanced_model and cuda and diffusion' --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: 4
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate

.buildkite/test-nightly.yml

@@ -12,7 +12,7 @@ steps:
     if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "omni-test"
     steps:
       - label: ":full_moon: Omni · Function Test with H100"
-        timeout_in_minutes: 90
+        timeout_in_minutes: 120
         commands:
           - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
         agents:
@@ -49,25 +49,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Omni · Function Test with L4"
-        timeout_in_minutes: 90
-        commands:
-          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
-        agents:
-          queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
       - label: ":full_moon: Omni · Doc Test with L4"
         timeout_in_minutes: 90
         commands:
@@ -172,7 +153,7 @@ steps:
 
       - label: ":full_moon: Omni · Perf Test · No Async Chunk"
         key: nightly-omni-performance-no-async-chunk
-        timeout_in_minutes: 180
+        timeout_in_minutes: 300
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
@@ -395,9 +376,11 @@ steps:
     if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
     steps:
       - label: ":full_moon: TTS · Function Test with L4"
-        timeout_in_minutes: 90
+        timeout_in_minutes: 120
         commands:
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - export VLLM_USE_DEEP_GEMM="0"
+          - export VLLM_MOE_USE_DEEP_GEMM="0"
           - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_voxcpm2_tts_expansion.py
         agents:
           queue: "gpu_1_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
@@ -568,54 +551,84 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-  - label: ":full_moon: TTS · Higgs-Audio-V3 · Perf Test"
-    key: nightly-tts-performance-higgs-audio-v3
-    depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
-    timeout_in_minutes: 180
-    commands:
-      - export BENCHMARK_DIR=tests/dfx/perf/results
-      - export VLLM_USE_DEEP_GEMM="0"
-      - export VLLM_MOE_USE_DEEP_GEMM="0"
-      - |
-        set +e
-        pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_higgs_audio_v3.json
-        EXIT=$$?
-        buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
-        exit $$EXIT
-    agents:
-      queue: "mithril-h100-pool"
-    plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                resources:
-                  limits:
-                    nvidia.com/gpu: 1
-                volumeMounts:
+      - label: ":full_moon: TTS · Higgs-Audio-V3 · Perf Test"
+        key: nightly-tts-performance-higgs-audio-v3
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - export VLLM_USE_DEEP_GEMM="0"
+          - export VLLM_MOE_USE_DEEP_GEMM="0"
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_higgs_audio_v3.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
                   - name: devshm
-                    mountPath: /dev/shm
+                    emptyDir:
+                      medium: Memory
                   - name: hf-cache
-                    mountPath: /root/.cache/huggingface
-                env:
-                  - name: HF_HOME
-                    value: /root/.cache/huggingface
-                  - name: HF_TOKEN
-                    valueFrom:
-                      secretKeyRef:
-                        name: hf-token-secret
-                        key: token
-            nodeSelector:
-              node.kubernetes.io/instance-type: gpu-h100-sxm
-            volumes:
-              - name: devshm
-                emptyDir:
-                  medium: Memory
-              - name: hf-cache
-                hostPath:
-                  path: /mnt/hf-cache
-                  type: DirectoryOrCreate
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+  - group: ":card_index_dividers: Tiny Model Tests [Multi-GPU]"
+    key: nightly-tiny-model-test-group
+    depends_on: upload-nightly-pipeline
+    if: >-
+      build.env("NIGHTLY") == "1" ||
+      build.pull_request.labels includes "nightly-test"
+    steps:
+      - label: ":full_moon: Tiny Model Tests · Diffusion (Multi-GPU accelerations)"
+        timeout_in_minutes: 30
+        # NOTE: Single-GPU (core_model) tiny model tests run on the ready label,
+        # so nightly only adds the multi-GPU parallelism configurations.
+        #
+        # This is a bit confusing, but `full_model` is used to mark the multigpu tests, and
+        # --run-level `core_model` is used to indicate that we should use tiny model weights
+        # for the CI, since these tests should pass irrespective of whether or not we use
+        # the full weights or not.
+        commands:
+          - timeout 25m pytest -sv tests/model_tests/diffusion -m 'full_model and cuda' --run-level "core_model"
+        agents:
+          queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
   # Diffusion X2I suite: x2i / x2a / x2t and related non-video paths; x2v is only in "Diffusion X2V Model Test" below.
   - group: ":card_index_dividers: Diffusion X2I(&A&T) Model Test"
@@ -636,11 +649,7 @@ steps:
             tests/e2e/online_serving/test_qwen_image_expansion.py
             tests/e2e/online_serving/test_qwen_image_edit_expansion.py
             tests/e2e/online_serving/test_qwen_image_layered_expansion.py
-            tests/e2e/online_serving/test_longcat_image_expansion.py
-            tests/e2e/online_serving/test_longcat_image_edit_expansion.py
-            tests/e2e/online_serving/test_flux_2_dev_expansion.py
             tests/e2e/online_serving/test_bagel_expansion.py
-            tests/e2e/offline_inference/test_sensenova_u1_expansion.py
             -m "full_model and diffusion and H100 and not distributed_cuda"
             --run-level "full_model"
         agents:
@@ -726,10 +735,6 @@ steps:
           - >-
             pytest -sv
             tests/e2e/online_serving/test_qwen_image_layered_expansion.py
-            tests/e2e/online_serving/test_longcat_image_expansion.py
-            tests/e2e/online_serving/test_longcat_image_edit_expansion.py
-            tests/e2e/online_serving/test_flux_2_dev_expansion.py
-            tests/e2e/offline_inference/test_glm_image_autoround_w4a16_expansion.py
             -m "full_model and diffusion and H100 and distributed_cuda"
             --run-level "full_model"
         agents:
@@ -808,13 +813,14 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU DreamZero"
+      - label: ":full_moon: Diffusion X2I(&A&T) · GLM-Image Function Test with H100"
         timeout_in_minutes: 120
         commands:
           - >-
             pytest -sv
-            tests/e2e/online_serving/test_dreamzero_expansion.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
+            tests/e2e/online_serving/test_glm_image_expansion.py
+            tests/e2e/offline_inference/test_glm_image_autoround_w4a16_expansion.py
+            -m "full_model and diffusion and H100"
             --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
@@ -851,47 +857,8 @@ steps:
                       type: DirectoryOrCreate
 
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · GLM-Image Function Test with H100"
-        timeout_in_minutes: 120
-        commands:
-          - pytest -sv tests/e2e/online_serving/test_glm_image_expansion.py -m "full_model and diffusion and H100" --run-level "full_model"
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 2
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
-
       - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
-        timeout_in_minutes: 60
+        timeout_in_minutes: 120
         commands:
           - pytest -sv tests/e2e/ -k "not test_wan and not test_bagel_expansion and not hunyuan" -m "full_model and diffusion and L4" --run-level "full_model" --ignore=tests/e2e/accuracy
         agents:
@@ -1276,10 +1243,10 @@ steps:
       build.pull_request.labels includes "nightly-test" ||
       build.pull_request.labels includes "diffusion-x2v-test"
     steps:
-      - label: ":full_moon: Diffusion X2V · Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test"
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_hunyuan_video_15_expansion.py tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py tests/e2e/online_serving/test_ltx2_expansion.py -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda" --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -1314,14 +1281,166 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Diffusion X2V · Accuracy Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda" --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: ":full_moon: Diffusion X2V · Doc Test"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -s -v tests/examples/offline_inference/test_image_to_video.py -m "full_model and example and H100" --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: ":full_moon: Diffusion X2V · Other Function Test"
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_ltx2_expansion.py -m "full_model and cuda" --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: ":full_moon: Diffusion X2V · Wan2.2 Accuracy Test"
         timeout_in_minutes: 180
         commands:
           - pytest -s -v tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py -m full_model --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2v_matches_diffusers' -m full_model --run-level full_model
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: ":full_moon: Diffusion X2V · Other Accuracy Test"
+        timeout_in_minutes: 180
+        commands:
           - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_t2v/test_hunyuanvideo15_t2v_video_similarity.py -m full_model --run-level full_model
           - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py -m full_model --run-level full_model
           - pytest -s -v tests/e2e/accuracy/test_ltx2_3_video_similarity.py -m full_model --run-level full_model
-          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2v_matches_diffusers' -m full_model --run-level full_model
           - pytest -s -v tests/e2e/accuracy/test_video_streaming_output_similarity.py -m full_model --run-level full_model
         agents:
           queue: "mithril-h100-pool"
@@ -1466,7 +1585,7 @@ steps:
     key: nightly-testcase-statistics
     timeout_in_minutes: 120
     depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
+    if: build.env("NIGHTLY") == "1"
     commands:
       - python tools/nightly/buildkite_testcase_statistics.py -o tests/dfx/perf/results/buildkite_testcase_statistics.html
       - buildkite-agent artifact upload "tests/dfx/perf/results/*.html"

.buildkite/test-ready.yml

@@ -157,6 +157,27 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
 
+  - group: ":card_index_dividers: Tiny Model Tests [Single GPU]"
+    depends_on: upload-ready-pipeline
+    steps:
+      - label: "Tiny Model Tests · Diffusion (base)"
+        commands:
+          - timeout 15m pytest -sv tests/model_tests/diffusion -m 'core_model and cuda' --run-level "core_model"
+        agents:
+          # Ready label only runs the cases that don't need parallelism for tiny
+          # model tests, which is usually just a smoke test for TTI and I2I
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
   - label: "Engine&Model Executor Test"
     depends_on: upload-ready-pipeline
     commands:
@@ -184,96 +205,43 @@ steps:
     commands:
       - |
         timeout 20m bash -c "
-                  pytest -sv tests/distributed/ -m 'core_model and cuda' --run-level 'core_model'
+          pytest -sv tests/distributed/ -m 'core_model and cuda and L4' --run-level 'core_model'
         "
     agents:
-      queue: "mithril-h100-pool"
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
     plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$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
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+            - "HF_TOKEN"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - group: ":card_index_dividers: Entrypoints Test"
+  - label: "Entrypoints Test"
     depends_on: upload-ready-pipeline
-    steps:
-      - label: "Entrypoints Test with H100"
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -sv tests/entrypoints/ -m 'core_model and cuda and H100' --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:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 2
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
-      - label: "Entrypoints Test with L4"
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -sv tests/entrypoints/ -m 'core_model and cuda and L4' --run-level 'core_model'
-            "
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
+    commands:
+      - |
+        timeout 30m bash -c "
+          export VLLM_WORKER_MULTIPROC_METHOD=spawn
+          pytest -sv tests/entrypoints/ -m 'core_model and cuda' --run-level 'core_model'
+        "
+    agents:
+      queue: "gpu_1_queue"
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+            - "HF_TOKEN"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
 
   - group: ":card_index_dividers: E2E Test"
     depends_on: upload-ready-pipeline
@@ -584,6 +552,50 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+      - label: "Diffusion · Cosmos3 Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_cosmos3.py
+          - vllm_omni/diffusion/models/cosmos3/
+        commands:
+          - |
+            timeout 40m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_cosmos3.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
 
       - label: "Diffusion · Qwen Image Test"
         source_file_dependencies:

.buildkite/test-weekly.yml

@@ -206,3 +206,246 @@ steps:
                 - "HF_TOKEN"
               volumes:
                 - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - group: ":card_index_dividers: E2E Tests"
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    steps:
+      - label: "Omni · L4"
+        timeout_in_minutes: 90
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
+          - >-
+            pytest -s -v tests/e2e/
+            --ignore=tests/e2e/online_serving/test_qwen2_5_omni_expansion.py
+            --ignore=tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py
+            -m "slow and L4 and omni"
+            --run-level "full_model"
+        agents:
+          queue: "gpu_4_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "Omni · H100"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and H100 and omni"
+            --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: "TTS · L4"
+        timeout_in_minutes: 90
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - export VLLM_USE_DEEP_GEMM="0"
+          - export VLLM_MOE_USE_DEEP_GEMM="0"
+          - |
+            failed=0
+            for f in $(pytest --collect-only -q tests/e2e/ -m "slow and L4 and tts" --run-level "full_model" | grep '\.py::' | sed -E 's|::.*||' | sort -u); do
+              pytest -s -v "$$f" -m "slow and L4 and tts" --run-level "full_model" || failed=1
+            done
+            exit $$failed
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "TTS · H100"
+        timeout_in_minutes: 90
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - export VLLM_USE_DEEP_GEMM="0"
+          - export VLLM_MOE_USE_DEEP_GEMM="0"
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and H100 and tts"
+            --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: "Diffusion · L4"
+        timeout_in_minutes: 60
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and L4"
+            --run-level "full_model"
+        agents:
+          queue: "gpu_4_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "Diffusion · H100 · Single-GPU"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and H100 and not distributed_cuda"
+            --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: "Diffusion · H100 · Multi-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and H100 and distributed_cuda"
+            --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate

.github/CODEOWNERS

@@ -22,30 +22,30 @@
 /.buildkite/ @yenuo26 @congw729
 
 # Engine and model_executor
-/vllm_omni/core/ @tzhouam @yinpeiqi @fake0fan
-/vllm_omni/engine/ @tzhouam @yinpeiqi @fake0fan
-/vllm_omni/entrypoints/ @tzhouam @yinpeiqi @fake0fan
+/vllm_omni/core/ @tzhouam @yinpeiqi @fake0fan @Sy0307
+/vllm_omni/engine/ @tzhouam @yinpeiqi @fake0fan @Sy0307
+/vllm_omni/entrypoints/ @tzhouam @yinpeiqi @fake0fan @alex-jw-brooks
 /vllm_omni/model_executor/ @tzhouam @gcanlin
 /vllm_omni/worker/ @tzhouam @gcanlin
 
 # Diffusion
-/vllm_omni/diffusion/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @david6666666
+/vllm_omni/diffusion/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @david6666666 @xuechendi
 /vllm_omni/lora/ @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @yuanheng-zhao
 
 # Diffusion CPU offloading
-/vllm_omni/diffusion/offloader/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @yuanheng-zhao
+/vllm_omni/diffusion/offloader/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @yuanheng-zhao @xuechendi
 
 # Omni and TTS support
 /vllm_omni/transformers_utils/ @tzhouam @yuanheng-zhao
-/vllm_omni/model_executor/models/ @linyueqian @ZeldaHuang @princepride @yuanheng-zhao @amy-why-3459 @Bounty-hunter
+/vllm_omni/model_executor/models/ @linyueqian @ZeldaHuang @princepride @yuanheng-zhao @amy-why-3459 @Bounty-hunter @Sy0307 @gcanlin
 
 # Benchmarks
 /benchmarks/accuracy/ @david6666666
 /benchmarks/build_dataset/ @linyueqian @yuanheng-zhao
-/benchmarks/diffusion/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa
+/benchmarks/diffusion/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @xuechendi
 /benchmarks/distributed/ @princepride
 /benchmarks/glm_image/ @princepride @JaredforReal
-/benchmarks/tts/ @linyueqian @ZeldaHuang @princepride
+/benchmarks/tts/ @linyueqian @ZeldaHuang @princepride @Sy0307 @gcanlin
 
 # Usability and Observability
 /vllm_omni/deploy/ @tzhouam @lishunyang12
@@ -53,13 +53,13 @@
 /vllm_omni/profiler/ @gcanlin
 
 # ComfyUI
-/apps/ @fhfuih
+/apps/ @fhfuih @alex-jw-brooks
 
 # Large Scale Serving
 /vllm_omni/distributed/ @princepride @yuanheng-zhao @xuechendi @natureofnature @fake0fan
 
 # Quantization and configuration
-/vllm_omni/config/ @lishunyang12
+/vllm_omni/config/ @lishunyang12 @alex-jw-brooks
 /vllm_omni/quantization/ @david6666666 @Isotr0py @lishunyang12
 
 # Hardware plugins and accelerators

.gitignore

@@ -208,6 +208,7 @@ cache/
 .cache/
 diffusion_cache/
 kv_cache/
+!vllm_omni/experimental/ar_diffusion/kv_cache/
 
 # Logs
 logs/
@@ -274,3 +275,8 @@ tests/.ci_generated/
 
 # Local Lance precision-matching dump scripts (not part of the repo)
 lance_compare/
+
+# FP8 quantization test sample images: tests/diffusion/quantization/
+# test_quantization_fp8.py saves these into the repo root for manual
+# inspection; they are byproducts and must never be committed.
+test_*_fp8.png

benchmarks/fish-speech/bench_speaker_cache.py

@@ -201,6 +201,7 @@ async def main():
             "input": prompt,
             "voice": "default",
             "stream": True,
+            "stream_format": "audio",
             "response_format": "pcm",
             "ref_audio": ref_audio_b64,
             "ref_text": args.ref_text,
@@ -240,6 +241,7 @@ async def main():
             "input": prompt,
             "voice": args.voice_name,
             "stream": True,
+            "stream_format": "audio",
             "response_format": "pcm",
             "ref_text": args.ref_text,
             "max_new_tokens": 2048,

docker/Dockerfile.ci

@@ -1,8 +1,8 @@
-# Target vllm version: v0.21.0 (releases/v0.21.0 branch).
+# Target vllm version: v0.24.0 (releases/v0.24.0 branch).
 # The pinned VLLM_PRECOMPILED_WHEEL_COMMIT and VLLM_BASE_TAG below must
 # be updated on every rebase to match the new upstream vllm commit/tag.
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.23.0
+ARG VLLM_BASE_TAG=v0.24.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}
@@ -14,8 +14,71 @@ RUN apt-get update && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
+
+# Install vLLM (CUDA 13.0 build) from precompiled wheel at the selected commit.
+# Must use direct URL because the wheel has a PEP 440 local version identifier
+# (e.g. +gaeee7ef93.cu130) which pip/uv refuse to install from a PEP 503 package index.
+# The cu130 wheel index lives at /<commit>/cu130/vllm/ but the actual wheel files
+# are stored at the top level /<commit>/<wheel>.
+ENV VLLM_PRECOMPILED_WHEEL_COMMIT=dd10e03f95f94edbea1975c67ace3a35ec9a8a40
+# Keep the base image's CUDA-12 toolkit libs on the loader path. NOTE: this does
+# NOT resolve libnvJitLink.so.13 (that CUDA-13 lib is not in /usr/local/cuda here) —
+# the .so.13 JIT linker is handled separately via ldconfig after the cu130 wheels
+# are installed (see the nvidia-cuda-nvjitlink-cu13 step below).
+ENV LD_LIBRARY_PATH=/usr/local/cuda/targets/x86_64-linux/lib:${LD_LIBRARY_PATH}
+RUN VLLM_WHEEL_URL=$(python3 -c "import urllib.request,re; \
+    from urllib.parse import urljoin, quote; \
+    index_url='https://wheels.vllm.ai/${VLLM_PRECOMPILED_WHEEL_COMMIT}/cu130/vllm/'; \
+    html=urllib.request.urlopen(index_url).read().decode(); \
+    m=re.search(r'href=\"([^\"]*x86_64\.whl)\"',html); \
+    assert m, f'No x86_64 vLLM wheel found at {index_url}'; \
+    print(quote(urljoin(index_url,m.group(1)), safe=':/+%'))") && \
+    echo "Installing vLLM from: ${VLLM_WHEEL_URL}" && \
+    uv pip install --system --force-reinstall "${VLLM_WHEEL_URL}"
+
+
+
 RUN uv pip install --system ".[dev]"
 
+RUN uv pip install --system --upgrade \
+        "flashinfer-python==0.6.13" \
+        "flashinfer-cubin==0.6.13"
+
+RUN uv pip install --system --upgrade \
+    "flashinfer-jit-cache==0.6.13" \
+    --index-url https://flashinfer.ai/whl/cu130
+
+# Force-reinstall torch AFTER flashinfer, because flashinfer --upgrade pulls in
+# a newer torch (2.12.x) whose libtorch_cuda.so lacks
+# at::cuda::getCurrentCUDABlasHandle, causing "undefined symbol" at import time.
+RUN uv pip install --system --force-reinstall \
+    "torch==2.11.0" "torchaudio==2.11.0" "torchvision==0.26.0" \
+    --torch-backend cu130
+
+# CUDA 13 JIT linker: libnvJitLink.so.13 is provided by the nvidia-nvjitlink
+# package that the cu130 torch/vllm wheels above pull into site-packages (the
+# v0.24.0 base image's /usr/local/cuda toolkit is CUDA 12, so the LD_LIBRARY_PATH
+# above never resolves the .so.13). Its lib dir is not on the dynamic-loader path,
+# so bitsandbytes / torch.compile worker subprocesses fail to dlopen it. Locate the
+# already-installed lib and register its dir with ldconfig (system-wide, survives
+# subprocess env resets). Only fall back to the pip wheel (correct name:
+# nvidia-nvjitlink-cu13, NOT nvidia-cuda-nvjitlink-cu13, which does not exist) if
+# the torch deps did not bring it in.
+RUN NVJITLINK_LIB="$(find / -name 'libnvJitLink.so.13*' 2>/dev/null | head -1)" && \
+    if [ -z "$NVJITLINK_LIB" ]; then \
+        uv pip install --system "nvidia-nvjitlink-cu13" && \
+        NVJITLINK_LIB="$(find / -name 'libnvJitLink.so.13*' 2>/dev/null | head -1)"; \
+    fi && \
+    test -n "$NVJITLINK_LIB" && \
+    dirname "$NVJITLINK_LIB" > /etc/ld.so.conf.d/nvjitlink-cu13.conf && \
+    ldconfig && \
+    echo "Registered ${NVJITLINK_LIB} with ldconfig"
+
+# Re-pin numpy LAST. numba requires numpy <= 2.4, but the flashinfer and torch cu130
+# reinstalls above resolve numpy back up to 2.5, which breaks `import numba` at test
+# time ("Numba needs NumPy 2.4 or less"). This must be the final dependency op so
+# nothing can clobber it again.
+RUN uv pip install --system "numpy==2.2.6"
 
 RUN ln -sf /usr/bin/python3 /usr/bin/python
 

docker/Dockerfile.cuda

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai:v0.23.0
+ARG BASE_IMAGE=vllm/vllm-openai:v0.24.0
 FROM ${BASE_IMAGE}
 
 ARG COMMON_WORKDIR=/app

docker/Dockerfile.npu

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.23.0-4885d8e
+ARG VLLM_ASCEND_TAG=v0.24.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 # WORKDIR /vllm-workspace/vllm

docker/Dockerfile.npu.a3

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.23.0-4885d8e-a3
+ARG VLLM_ASCEND_TAG=v0.24.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 # WORKDIR /vllm-workspace/vllm

docker/Dockerfile.npu.ci

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.23.0-4885d8e
+ARG VLLM_ASCEND_TAG=v0.24.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 ARG APP_DIR=/vllm-workspace/vllm-omni

docker/Dockerfile.npu.ci.a3

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.23.0-4885d8e-a3
+ARG VLLM_ASCEND_TAG=v0.24.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 ARG APP_DIR=/vllm-workspace/vllm-omni

docker/Dockerfile.rocm

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.23.0
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.24.0
 FROM ${BASE_IMAGE} AS base
 
 # Declare a variable to know if we want to use the nightly build or the stable build.

docker/Dockerfile.xpu

@@ -88,7 +88,7 @@ ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE="copy"
 
-ARG VLLM_VERSION=v0.23.0
+ARG VLLM_VERSION=v0.24.0
 RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
 WORKDIR /workspace/vllm
 
@@ -107,7 +107,9 @@ 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 .
+    uv pip uninstall triton triton-xpu && \
+    uv pip install triton-xpu==3.7.1 && \
+    uv pip install --no-build-isolation --no-deps .
 
 CMD ["/bin/bash"]
 
@@ -143,12 +145,11 @@ COPY . .
 
 ENV VLLM_OMNI_TARGET_DEVICE=xpu
 RUN --mount=type=cache,target=/root/.cache/uv \
-    echo "triton @ https://example.com/blocked" > /tmp/vllm_omni_constraints.txt && \
-    PIP_CONSTRAINT=/tmp/vllm_omni_constraints.txt \
     uv pip install --no-build-isolation \
     --extra-index-url https://mirrors.ustc.edu.cn/pypi/simple/ \
     ".[dev]" && \
-    uv pip uninstall oneccl oneccl-devel
+    uv pip uninstall triton oneccl oneccl-devel && \
+    uv pip install triton-xpu==3.7.1 --reinstall
 
 FROM vllm-omni AS vllm-omni-openai
 

docs/.nav.yml

@@ -68,10 +68,11 @@ nav:
         - Hybrid Sharded Data Parallel: user_guide/diffusion/parallelism/hsdp.md
         - Sequence Parallel: user_guide/diffusion/parallelism/sequence_parallel.md
         - Tensor Parallel: user_guide/diffusion/parallelism/tensor_parallel.md
-        - VAE Patch Parallel: user_guide/diffusion/parallelism/vae_patch_parallel.md
+        - VAE Parallelism: user_guide/diffusion/parallelism/vae_parallelism.md
       - CPU Offloading: user_guide/diffusion/cpu_offload_diffusion.md
       - LoRA: user_guide/diffusion/lora.md
       - Custom Pipeline: features/custom_pipeline.md
+      - Request-Level Batching: user_guide/diffusion/request_batching.md
       - Step Execution: user_guide/diffusion/step_execution.md
     - Quantization:
       - Overview: user_guide/quantization/overview.md
@@ -118,6 +119,7 @@ nav:
       - design/feature/async_chunk.md
       - design/feature/vae_parallel.md
       - design/feature/diffusion_step_execution.md
+      - design/feature/diffusion_request_level_batching.md
       - design/feature/diffusion_continuous_batching.md
     - Module Design:
       - design/module/ar_module.md

docs/community/governance.md

@@ -33,16 +33,20 @@ Lead maintainers are responsible for the overall direction and strategy of the p
 
 Committers have write access and merge rights. They typically have deep expertise in specific areas of this project and shepherd the community contributions:
 
+- [@alex-jw-brooks](https://github.com/alex-jw-brooks): Frontend, Configuration, and Entrypoint Support
+- [@amy-why-3459](https://github.com/amy-why-3459): Omni Modality Model and World Model Support
 - [@david6666666](https://github.com/david6666666): Quantization and Community Relationship
-- [@gcanlin](https://github.com/gcanlin): Hardware plugin and NPU integration
+- [@gcanlin](https://github.com/gcanlin): Hardware plugin, NPU integration, and TTS Model Support
 - [@Isotr0py](https://github.com/Isotr0py): Diffusion and Quantization
 - [@linyueqian](https://github.com/linyueqian): TTS and Omni Support
 - [@lishunyang12](https://github.com/lishunyang12): Quantization and Configuration
 - [@princepride](https://github.com/princepride): Diffusion and Omni Support
 - [@RuixiangMa](https://github.com/RuixiangMa): Diffusion models, parallel, cache, and docs
 - [@SamitHuang](https://github.com/SamitHuang): RL, Diffusion, and cache
+- [@Sy0307](https://github.com/Sy0307): TTS Model and AR Engine Support
 - [@tzhouam](https://github.com/tzhouam): Engine and New Model Support
 - [@wtomin](https://github.com/wtomin): Diffusion models, parallel, and docs
+- [@xuechendi](https://github.com/xuechendi): Diffusion and Intel GPU Support
 - [@ZeldaHuang](https://github.com/ZeldaHuang): Omni Support
 - [@ZJY0516](https://github.com/ZJY0516): Diffusion attention backend, kernel fusion, and CustomOp
 - [@yuanheng-zhao](https://github.com/yuanheng-zhao): Diffusion cache, offload, and Omni Support
@@ -53,7 +57,7 @@ This section breaks down diffusion-related responsibilities. If you have PRs tou
 
 ### Runtime-related
 
-- Diffusion models: @RuixiangMa, @wtomin
+- Diffusion models: @RuixiangMa, @wtomin, @xuechendi
 
 ### Optimization
 
@@ -61,6 +65,7 @@ This section breaks down diffusion-related responsibilities. If you have PRs tou
 - Attention backend: @ZJY0516
 - Cache: @yuanheng-zhao, @RuixiangMa, @SamitHuang
 - Offload: @yuanheng-zhao
+- Intel GPU support: @xuechendi
 - Quantization: @lishunyang12, @david6666666
 - Kernel fusion / communication-computation: @ZJY0516
 

docs/configuration/composable_parallel.md

@@ -0,0 +1,239 @@
+# Composable parallel strategies
+
+In vLLM-Omni, the `composable_parallel` system layers a declarative *parallel strategy* on top of the existing stage configs. A strategy YAML maps each pipeline stage (by its `model_stage` name) to a stack of axis declarations — one per parallelism dimension — that the engine translates into per-stage parallel sizing (`tensor_parallel_size`, `data_parallel_size`, `pipeline_parallel_size`, expert-parallel toggle, replica count) and, when relevant, an orchestrator-wide load-balancer policy.
+
+Strategies are intentionally narrow: each one only writes the axes it declares, leaving every other engine arg untouched. This makes a strategy file a small, reusable overlay you can swap independently of the deploy YAML (see `configuration/stage_configs.md`).
+
+!!! note
+    Composable parallel is **opt-in**. Passing `--strategy-config <path/to/strategy.yaml>` overlays the derived sizing onto the registry-merged stages *before* any CLI overrides are applied. When `--strategy-config` is omitted, stages remain exactly as their deploy YAML produces them.
+
+## Top-level schema reference
+
+A strategy file has a single top-level mapping, `strategies`, keyed by the `model_stage` name of each stage you want to overlay. Each value is a list of axis declarations (see [StrategySpec fields](#strategyspec-fields) below).
+
+| Field | Type | Required | Default | Description |
+|-------|------|----------|---------|-------------|
+| `strategies` | `dict[str, list]` | required | — | Maps `model_stage` name (e.g. `thinker`, `talker`, `code2wav`, `dit`) to a list of axis declarations. The key MUST be the stage name, not the integer `stage_id`. See `vllm_omni/config/composable_parallel/strategy_loader.py`. |
+
+### StrategySpec fields
+
+Each entry in a stage's list becomes one `StrategySpec` (see `vllm_omni/config/composable_parallel/spec.py`). The loader fills in `routing` and `aggregation` from per-kind defaults, so the file stays compact.
+
+| Field | Type | Required | Default | Description |
+|-------|------|----------|---------|-------------|
+| `axis` | str | required | — | Mesh-axis kind. One of `tp`, `dp`, `pp`, `ep`, `stage_replica` (wired today) or `sp_ulysses`, `sp_ring`, `cfg`, `vae_pp`, `hsdp`, `stage_pp`, `cp` (reserved; declaring one raises `AxisTranslationError` at translate time). See `MeshAxisKind` in `vllm_omni/config/composable_parallel/spec.py`. |
+| `size` | int | required | — | Axis degree. Must be `> 0`. |
+| `routing` | str | optional | per-kind default | Stateless routing policy. Accepted only for `dp` / `stage_replica`; other kinds reject it. Values: `random`, `round_robin`, `least_queue`. (`hash` is recognised but currently raises `NotImplementedError`.) See `RouteByStage` in `vllm_omni/config/composable_parallel/routing.py`. |
+| `l1_owner` | str | optional | per-kind default | Overrides which layer owns request routing for this axis. `engine` for `tp`/`dp`/`pp`/`ep`, `delegated` for `stage_replica`. Recorded into `shard_extension`. |
+| `name` | str | optional | axis kind | Human-readable label for this spec, used in error messages. |
+
+Two additional `StrategySpec` slots — `layer_hook_specs` and `kernel_specs` — let advanced authors bind layer-walker hook IDs (`LayerHookSpec`) or kernel IDs (`KernelSpec`) onto an axis. These are not exposed by the YAML loader today and are intended for programmatic construction. See `vllm_omni/config/composable_parallel/spec.py` for the dataclasses.
+
+### CLI flags
+
+| Flag | Description |
+|------|-------------|
+| `--strategy-config PATH` | Path to a `strategy.yaml`. Only takes effect on registry-based deploy paths (`--deploy-config` or the bundled default deploy YAML). Its derived sizing is overlaid onto the merged stages *before* CLI overrides. |
+| `--stage-configs-path PATH` | **Legacy.** Loads the old `stage_args` YAML schema for un-migrated models. Passing `--strategy-config` together with this path is silently inapplicable — vLLM-Omni emits a warning and the strategy is ignored (see `vllm_omni/entrypoints/utils.py`, lines 372–380). |
+| `--omni-lb-policy POLICY` | Orchestrator-wide load-balancer policy. Strategy-derived from any `stage_replica` axis; an explicit CLI value wins and must match the derived one or `AsyncOmniEngine` raises `Conflicting load-balancer policy` at construction. |
+
+!!! warning
+    `--strategy-config` cannot be used with `--stage-configs-path` (the legacy deploy path). If your model has not been migrated to the registry-based pipeline yet, the strategy is silently dropped with a warning. Migrate to a registry-based model or use `--deploy-config` to apply a strategy.
+
+## Sub-schemas
+
+### `axis`
+
+The set of recognised axis kinds is the `MeshAxisKind` enum in `vllm_omni/config/composable_parallel/spec.py`:
+
+```yaml
+# Translatable today (see `_SUPPORTED_KINDS` in translator.py).
+axis: tp              # tensor parallel — sets tensor_parallel_size
+axis: dp              # engine data parallel — sets data_parallel_size (vLLM's intra-engine DP LB)
+axis: pp              # pipeline parallel — sets pipeline_parallel_size
+axis: ep              # dense expert parallel — sets enable_expert_parallel=True; size must equal tp*dp
+axis: stage_replica   # omni-coordinator-level replica fan-out — sets num_replicas + omni_lb_policy
+
+# Reserved (raise AxisTranslationError today; declared in the type system for forward compat).
+axis: sp_ulysses
+axis: sp_ring
+axis: cfg
+axis: vae_pp
+axis: hsdp
+axis: stage_pp
+axis: cp
+```
+
+### `routing`
+
+The full closed hierarchy lives in `vllm_omni/config/composable_parallel/routing.py` (`Broadcast`, `PartitionByHash`, `ShardSequence`, `ShardSpatial`, `RouteByExpert`, `PipelineMicrobatch`, `DuplicateWithCondTag`, `RouteByStage`). For YAML authors only `RouteByStage` is exposed as a `routing:` string, and only for `dp` / `stage_replica` axes (other axes have fixed routing — `tp`/`ep` broadcast, `pp` microbatch). Recognised policies:
+
+```yaml
+routing: random         # uniform-random across replicas/DP ranks
+routing: round_robin    # cycle through replicas/DP ranks (the default for dp/stage_replica)
+routing: least_queue    # send to the replica with the smallest backlog
+# routing: hash         # recognised but raises NotImplementedError; no key-stable balancer yet
+```
+
+### `aggregation`
+
+Aggregation is filled in automatically by the loader (`_default_aggregation` in `vllm_omni/config/composable_parallel/strategy_loader.py`) from the axis kind, so users do not author it:
+
+| Axis kind | Aggregation pattern |
+|-----------|---------------------|
+| `tp`, `ep` | `TakeRank` (all ranks agree; pick one) |
+| `dp`, ep (in some flows) | `Union` (disjoint replica results) |
+| `pp` | `StitchPipeline` |
+| `stage_replica` | `FanInByStage` |
+
+Full pattern hierarchy: `TakeRank`, `Union`, `GatherDim`, `AllGather`, `StitchSpatial`, `StitchPipeline`, `Combine`, `FanInByStage` (see `vllm_omni/config/composable_parallel/aggregation.py`).
+
+## Precedence
+
+From highest to lowest, when the same axis is touched by multiple layers:
+
+1. **CLI overrides** (`--stage_N_tensor_parallel_size`, `--stage-overrides`, `--omni-lb-policy`, etc.). A CLI value always wins; a warning is emitted whenever it overrides a strategy-declared axis (see `_reconcile_strategy_with_cli` in `vllm_omni/config/config_factory.py`).
+2. **Strategy YAML** (`--strategy-config`). Derived sizing is written onto the merged stages with *conflict-on-explicit* semantics — if the deploy YAML already set a value to something different, application raises `StrategyApplyError`.
+3. **Deploy YAML** (`--deploy-config` or the bundled `vllm_omni/deploy/<model_type>.yaml`).
+4. **Parser defaults.**
+
+After CLI overrides land, the device-layout guard re-runs (`check_device_layout`, called from `_reconcile_strategy_with_cli`) against the *effective* `tp * dp * pp * num_replicas` and the resolved `devices` string, so a `--devices` or `--tensor-parallel-size` that conflicts with the strategy's world size cannot slip past the pre-spawn check.
+
+## Worked example
+
+### Example 1 — `strategy_tp2.yaml`
+
+Shard the `thinker` stage with TP=2 on top of the bundled Qwen2.5-Omni deploy YAML. Routing is fixed (broadcast) for TP, so no `routing:` field is needed.
+
+```yaml
+# examples/offline_inference/qwen2_5_omni/strategy_tp2.yaml
+strategies:
+  thinker:
+    - axis: tp
+      size: 2
+```
+
+Launch:
+
+```bash
+python3 end2end.py --query-type text --num-prompts 6 \
+    --strategy-config strategy_tp2.yaml \
+    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
+```
+
+Effective `OmniParallelConfig` per stage after the overlay (see `OmniParallelConfig` in `vllm_omni/config/composable_parallel/translator.py`):
+
+| Stage (`model_stage`) | `tensor_parallel_size` | `data_parallel_size` | `pipeline_parallel_size` | `stage_replica_size` | `omni_lb_policy` |
+|-----------------------|------------------------|----------------------|--------------------------|----------------------|------------------|
+| `thinker`             | **2** (from strategy)  | 1                    | 1                        | 1                    | —                |
+| `talker`              | 1                      | 1                    | 1                        | 1                    | —                |
+| `code2wav`            | 1                      | 1                    | 1                        | 1                    | —                |
+
+### Example 2 — `strategy_stage_replica.yaml`
+
+Replicate the `thinker` stage across two independent engines and let omni's StagePool round-robin requests across them.
+
+```yaml
+# examples/offline_inference/qwen2_5_omni/strategy_stage_replica.yaml
+strategies:
+  thinker:
+    - axis: stage_replica
+      size: 2
+      routing: round_robin
+```
+
+Launch:
+
+```bash
+python3 end2end.py --query-type text --num-prompts 6 \
+    --strategy-config strategy_stage_replica.yaml \
+    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
+```
+
+Effective per-stage config and the derived orchestrator knob:
+
+| Stage (`model_stage`) | `tensor_parallel_size` | `stage_replica_size` | `omni_lb_policy` (derived) |
+|-----------------------|------------------------|----------------------|----------------------------|
+| `thinker`             | 1                      | **2**                | **`round-robin`**          |
+| `talker`              | 1                      | 1                    | —                          |
+| `code2wav`            | 1                      | 1                    | —                          |
+
+The derived `omni_lb_policy = round-robin` is **not** written into any stage's engine args. Instead it is surfaced on `StrategyApplyResult.omni_lb_policy` and applied once at orchestrator construction by `AsyncOmniEngine._apply_strategy_lb_policy` (see `vllm_omni/engine/async_omni_engine.py`, lines 1107–1134). If you also pass `--omni-lb-policy round-robin` the values must match; passing a different value raises `Conflicting load-balancer policy`.
+
+## Per-field deep-dive
+
+### `name`
+
+Default: the axis kind (e.g. `"tp"`).
+
+Human-readable label that appears in `AxisTranslationError` and `StrategyApplyError` messages. Set it to disambiguate two stages declaring the same axis kind when scanning logs.
+
+### `mesh_axis` (`axis` + `size`)
+
+Default: no default — both fields are required.
+
+`axis` chooses the parallelism dimension; `size` is its degree. Five kinds are wired end-to-end today (`tp`, `dp`, `pp`, `ep`, `stage_replica`); the remaining kinds in `MeshAxisKind` are reserved and raise `AxisTranslationError` at translate time so a typo cannot silently become a no-op (see `_SUPPORTED_KINDS` in `vllm_omni/config/composable_parallel/translator.py`). For `ep`, the size MUST equal `tensor_parallel_size * data_parallel_size`.
+
+### `routing`
+
+Default: per-kind — `RouteByStage(round_robin)` for `dp` and `stage_replica`; `PipelineMicrobatch()` for `pp`; `Broadcast()` for `tp`/`ep` and any other kind.
+
+User-facing string only on `dp` and `stage_replica` axes. The four recognised values are `random`, `round_robin`, `least_queue`, `hash`; `hash` is rejected today with `NotImplementedError`. See `RouteByStage` in `vllm_omni/config/composable_parallel/routing.py`.
+
+### `l1_owner`
+
+Default: per-kind — `engine` for `tp`/`dp`/`pp`/`ep`, `delegated` for `stage_replica`.
+
+Overrides which layer owns request routing along this axis. `engine` means vLLM realizes routing intra-engine; `delegated` means the omni coordinator owns routing across independent replicas. Set this only when overriding the per-kind default in `_DEFAULT_L1_OWNER` (`vllm_omni/config/composable_parallel/translator.py`); invalid values fail fast in `_resolve_l1_owner`.
+
+### `stage_replica` and `omni_lb_policy`
+
+`stage_replica` is *not* a vLLM world dimension; it spins up `size` independent engine replicas of one stage, coordinated by the omni `StagePool` load balancer. The axis writes both:
+
+- The stage's `num_replicas` (a per-stage deploy knob).
+- The pipeline-wide `omni_lb_policy` string (an orchestrator-level knob).
+
+Because `omni_lb_policy` is read once at `AsyncOmniEngine` construction — not from per-stage engine args — only one value is allowed across the whole pipeline. If two stages declare conflicting policies, application raises (see `apply_strategy_specs` in `vllm_omni/config/composable_parallel/apply.py`, lines 253–260). To set this knob manually instead, use `--omni-lb-policy` on the CLI; see [the model stage docs](stage_configs.md) for how stage replicas plug into the orchestrator at runtime.
+
+### `model_stage` keys
+
+Default: no default — required for every entry under `strategies:`.
+
+Strategy keys MUST be `model_stage` *names* (strings), the same labels stage configs use for `engine_args.model_stage` (see [`model_stage` in `stage_configs.md`](stage_configs.md#engine_argsmodel_stage)). Integer `stage_id` values are not accepted; the resolver matches on `model_stage` exactly (`_resolve_stage` in `vllm_omni/config/composable_parallel/apply.py`).
+
+## Common pitfalls
+
+!!! warning "Using `stage_id` integers as keys"
+    Strategy keys are `model_stage` names (str), not integer stage IDs. A file that uses an integer index will raise at apply time:
+
+    ```text
+    StrategyApplyError: strategy role model_stage=0 did not match any stage;
+    available stages: 'thinker'(id=0), 'talker'(id=1), 'code2wav'(id=2)
+    ```
+
+    Use the `model_stage` name (e.g. `thinker`) instead.
+
+!!! warning "Mixing CLI overrides with a strategy on the same axis"
+    A strategy is meant to be the single writer for the axes it declares. If a CLI override (`--stage_0_tensor_parallel_size`, `--stage-overrides '{"0": {...}}'`) sets the same axis, the **CLI value wins** and a warning is emitted from `_reconcile_strategy_with_cli`. The device-layout guard then re-runs against the effective layout, so a CLI override that breaks `tp * dp * pp * num_replicas` will still fail fast at config time.
+
+!!! warning "Passing `--strategy-config` against the legacy `--stage-configs-path`"
+    Composable parallel only applies on the registry-based deploy path. If a model still resolves through the legacy `stage_args` YAML (loaded via `--stage-configs-path`), `--strategy-config` is silently inapplicable and emits:
+
+    ```text
+    --strategy-config (<path>) was provided but model '<model>' resolves via the
+    legacy stage_configs YAML path, which does not support composable-parallel
+    strategies; the strategy is ignored. Use a registry-based model to apply it.
+    ```
+
+    Migrate the model to the new deploy schema (see `configuration/stage_configs.md`) before applying a strategy.
+
+!!! warning "Conflicting `omni_lb_policy` across stages"
+    `omni_lb_policy` is a single pipeline-wide knob. If two `stage_replica` axes in different stages derive different policies, `apply_strategy_specs` raises:
+
+    ```text
+    StrategyApplyError: role 'talker' derives omni_lb_policy='least-queue-length'
+    but role 'thinker' already derived 'round-robin'. The load-balancer policy is
+    pipeline-wide; only one value is allowed.
+    ```
+
+    Make the policies match, drop one, or supply the value explicitly via `--omni-lb-policy`.

docs/configuration/pd_disaggregation.md

@@ -123,7 +123,7 @@ After inserting the extra thinker stage, renumber the remaining stages:
     runtime:
       devices: "2"
     engine_input_source: [1]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker
+    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_token_only
 
   - stage_id: 3
     runtime:
@@ -131,7 +131,7 @@ After inserting the extra thinker stage, renumber the remaining stages:
     engine_args:
       max_num_seqs: 1
     engine_input_source: [2]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav
+    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav_token_only
 ```
 
 Compared with the default Qwen3-Omni config:

docs/configuration/stage_configs.md

@@ -290,7 +290,6 @@ stage_args:
       enable_prefix_caching: false
       engine_output_type: latent
     engine_input_source: [0]
-    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen2_5_omni.thinker2talker
     default_sampling_params:
       temperature: 0.9
       top_p: 0.8
@@ -379,6 +378,20 @@ The maximum number of sequences for concurrent processing in this stage. For LLM
 
 Default: `1`
 
+#### `engine_args.request_batch_max_wait_ms`
+
+The maximum time, in milliseconds, that a diffusion request-mode stage may wait
+before the first `schedule()` of a new scheduler wave so compatible requests can
+accumulate for request-level batching. This only applies to diffusion pipelines
+that support request-level batching with `step_execution` disabled.
+
+Use this together with `max_num_seqs > 1` for bursty serving traffic. `0`
+disables admission waiting and preserves the lowest first-request latency.
+For diffusion request-level batching tuning, see
+[Request-Level Batching](../user_guide/diffusion/request_batching.md).
+
+Default: `0.0`
+
 ### `engine_args`
 
 Engine arguments for configuring the LLM engine, diffusion engine, or other engine types used by this stage.

docs/contributing/ci/CI_5levels.md

@@ -67,7 +67,7 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
     <tr>
       <td><strong>L2</strong><br>(E2E across models & GPU-required UT)</td>
       <td>Online (basic deployment scenarios):<br>dummy, normal inference function (output format, stream), some instance startup UT</td>
-      <td>High-priority models + online basic scenarios + request success validation</td>
+      <td>High-priority models + online basic scenarios; request success, non-empty output, format match (no Whisper/accuracy)</td>
       <td>High-priority features (using random lightweight models)</td>
       <td>High-priority interfaces (using random lightweight models)</td>
       <td><code>core_model and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
@@ -89,7 +89,7 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
     <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>High/medium-priority models with real weights + online/offline key scenarios + basic accuracy validation</td>
+      <td>High/medium-priority models + key online/offline scenarios; real weights, Whisper/similarity, preset voice gender, basic accuracy</td>
       <td>Medium-priority features (using random lightweight models)</td>
       <td>Medium-priority interfaces (using random lightweight models)</td>
       <td><code>advanced_model and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
@@ -113,9 +113,9 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
     <tr>
       <td><strong>L4</strong><br>(Perf & Integration & Accuracy)</td>
       <td>Online: full functional scenarios + performance test + doc test + accuracy test</td>
-      <td>High-priority models: function, performance, accuracy, and doc testing<br>Medium/low-priority models: function and doc testing</td>
-      <td>Low-priority features (using random lightweight models)</td>
-      <td>Low-priority interfaces (using random lightweight models)</td>
+      <td>High-priority models: function, performance, accuracy, and doc testing<br>Medium-priority models: function and doc testing</td>
+      <td>Low-priority features (using real weights)</td>
+      <td>Low-priority interfaces (using real weights)</td>
       <td><code>full_model and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
       <td>&lt;3 hour</td>
       <td>
@@ -144,7 +144,7 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
     <tr>
       <td><strong>L5</strong><br>(Stability & Reliability)</td>
       <td>Online: long-term stability test + reliability test</td>
-      <td>Long-term stability and reliability testing for high-priority models</td>
+      <td>Long-term stability and reliability testing for high-priority models<br>Low-priority models: function and doc testing</td>
       <td>/</td>
       <td>Invalid-parameter validation for high-priority interfaces</td>
       <td><code>slow and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
@@ -446,7 +446,11 @@ L2 level testing builds upon L1 by introducing GPU resources and verifying that
     -   ***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.
+    -   ***Focus***: Uses `dummy` weights (via deploy YAML patching at `core_model`) 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, and **cheap payload checks** at `--run-level core_model` (see below). Also includes some unit tests that require launching independent service instances.
+    -   ***L2 response validation (`core_model`)***: Implemented in `tests/helpers/assertions.py` and invoked by `OpenAIClientHandler` based on `--run-level`. At L2 we require **request success** plus minimal output sanity—not full accuracy:
+        -   **Speech / TTS** (`assert_audio_speech_response`): decoded audio must be present and non-empty (or exceed `min_audio_bytes` when set in `request_config`); `response_format` must match the returned content-type (e.g. `wav`, `pcm`). Whisper transcript similarity, PCM HNR, and preset-voice gender checks run only at L3+.
+        -   **Diffusion** (`assert_diffusion_response`): at least one non-empty image, video, or audio artifact. Resolution/frame-count parity and other deep checks run only at L3+.
+        -   **Omni multimodal** (`assert_omni_response`): L2 asserts successful completion; keyword, transcript, and cross-modal similarity checks run only at L3+.
     -   ***Characteristic***: Requires ***GPU*** resources to perform model computations.
 
 ### 1.3 Test Directory and Execution Files
@@ -555,6 +559,11 @@ L3 level testing executes after code is merged into the main branch. Its core pu
 -   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.
+-   ***L3 response validation (`advanced_model`)***: At `--run-level advanced_model`, `tests/helpers/assertions.py` adds semantic checks on top of L2 payload gates (see Chapter 1):
+    -   **Speech / TTS** (`assert_audio_speech_response`): Whisper transcript of returned audio vs `request_config["input"]` (cosine similarity &gt; 0.9 when input text is set); optional `min_audio_bytes` floor; PCM harmonic-to-noise ratio when `response_format` is `pcm`; **preset voice gender** via `_assert_preset_voice_gender_from_audio` when `voice` matches a known preset in `_PRESET_VOICE_GENDER_MAP` (pitch/F0-based classifier; skipped for unknown voices or `pcm` output).
+    -   **Omni multimodal** (`assert_omni_response`): non-empty text/audio outputs; `key_words` in transcript or text; text–audio similarity / containment; preset **`speaker`** gender check (same helper as TTS).
+    -   **Diffusion** (`assert_*_diffusion_response`): image/video dimension and frame-count parity with request parameters where configured.
+    -   **Accuracy suites**: pixel/video similarity and metric baselines under `/tests/e2e/accuracy/` (separate from inline helper assertions).
 
 ### 2.3 Test Directory and Execution Files
 
@@ -582,9 +591,9 @@ L3 level testing executes after code is merged into the main branch. Its core pu
 
     **Explanation**:
 
-    @pytest.mark.advanced_model: Marks the test as L3 merge level, indicating deep validation with real models. @pytest.mark.full_model: Marks L4 nightly-only suites (e.g. `test_*_expansion.py`, doc examples).
+    @pytest.mark.advanced_model: Marks the test as L3 merge level (`--run-level advanced_model`): real weights, Whisper/similarity/keyword checks, diffusion deep checks, and **preset voice/speaker gender** validation where applicable. @pytest.mark.full_model: Marks L4 nightly-only suites (e.g. `test_*_expansion.py`, doc examples).
 
-    @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.core_model: Marks the test as L1 or L2 level. At `--run-level core_model`, validation is limited to request success plus cheap payload checks (e.g. non-empty audio bytes, response format/content-type, non-empty diffusion outputs)—not Whisper, keyword, or accuracy gates. Deploy YAML may use `load_format: dummy` for fast PR feedback.
 
     @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.
 
@@ -680,7 +689,7 @@ L3 level testing executes after code is merged into the main branch. Its core pu
 
     **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_omni_request` and `send_diffusion_request` methods internally includes validation logic dynamically selected based on the --run-level parameter: core_model performs basic validation, while advanced_model and full_model perform deep validation.
+    **Implicit Validation**: `send_omni_request`, `send_audio_speech_request`, and `send_diffusion_request` call `assert_*_response` helpers with the session `--run-level`. At `core_model`, checks are limited to success plus cheap payload sanity (non-empty audio/media, format/content-type, optional `min_audio_bytes`). At `advanced_model` and `full_model`, deep validation adds Whisper transcripts, keyword/similarity, diffusion dimension checks, PCM HNR, etc.
 
     **Audio output debugging**: Deep validation may transcribe returned audio via `convert_audio_bytes_to_text` (Whisper). If an audio keyword or text–audio similarity assertion fails, set `VLLM_OMNI_KEEP_REQUEST_MEDIA=1` before running pytest to keep the intermediate WAV files for inspection (see [Test helper environment variables](#test-helper-environment-variables)).
 

docs/contributing/model/adding_diffusion_model.md

@@ -349,12 +349,14 @@ class YourModelPipeline(nn.Module):
 - def __call__(
 + def forward(
     self,
-+   req: OmniDiffusionRequest,  # ← Add request parameter here
++   req: DiffusionRequestBatch,  # ← Add request-batch parameter here
 - ):
-+ ) -> DiffusionOutput:  # ← Add return type
++ ) -> list[DiffusionOutput]:  # ← Add return type
 ```
 
-[`OmniDiffusionRequest`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/request/#vllm_omni.diffusion.request.OmniDiffusionRequest) is a dataclass that contains the **prompts** and **sampling parameters** [`OmniDiffusionSamplingParams`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/inputs/data/#vllm_omni.inputs.data.OmniDiffusionSamplingParams) for the diffusion pipeline execution. It also contains a request_id for other components to trace this request and its outputs.
+[`OmniDiffusionRequest`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/request/#vllm_omni.diffusion.request.OmniDiffusionRequest) is a dataclass that contains one **prompt** and the **sampling parameters** [`OmniDiffusionSamplingParams`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/inputs/data/#vllm_omni.inputs.data.OmniDiffusionSamplingParams) for one logical diffusion request. It also contains a request_id for other components to trace this request and its outputs. Before pipeline execution, the runner wraps one or more independent requests into `DiffusionRequestBatch`.
+
+[`DiffusionRequestBatch`](https://docs.vllm.ai/projects/vllm-omni/en/latest/api/vllm_omni/diffusion/worker/request_batch/#vllm_omni.diffusion.worker.request_batch.DiffusionRequestBatch) exposes compatibility properties such as `prompts`, `sampling_params`, and `request_id`. Pipelines that can execute the whole request batch in one forward pass should set `supports_request_batch = True`; other pipelines still receive a single-request batch and return a one-element output list.
 
 See some parameters in `OmniDiffusionSamplingParams` as follows:
 
@@ -367,35 +369,36 @@ See some parameters in `OmniDiffusionSamplingParams` as follows:
 **Extract parameters from request:**
 
 ```python
-from vllm_omni.diffusion.request import OmniDiffusionRequest
 from vllm_omni.diffusion.data import DiffusionOutput
+from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
 
 def forward(
     self,
-    req: OmniDiffusionRequest,
-) -> DiffusionOutput:
-    # Extract prompts from request
-    if req.prompts is not None:
-        prompt = [
-            p if isinstance(p, str) else (p.get("prompt") or "")
-            for p in req.prompts
-        ]
+    req: DiffusionRequestBatch,
+) -> list[DiffusionOutput]:
+    # Extract prompts from the request batch
+    prompts = [
+        p if isinstance(p, str) else (p.get("prompt") or "")
+        for p in req.prompts
+    ]
 
-    # Extract sampling parameters
+    # Extract common sampling parameters
     sampling_params = req.sampling_params
     num_inference_steps = sampling_params.num_inference_steps or 50
     guidance_scale = sampling_params.guidance_scale or 7.5
     height = sampling_params.height or (self.default_sample_size * self.vae_scale_factor)
     width = sampling_params.width or (self.default_sample_size * self.vae_scale_factor)
 
-    # For image editing pipelines, extract images from multi_modal_data
-    if hasattr(req, 'multi_modal_data') and req.multi_modal_data:
-        input_images = req.multi_modal_data.get('image', [])
+    # For image editing pipelines, extract media from each prompt dict
+    input_images = []
+    for p in req.prompts:
+        multi_modal_data = p.get("multi_modal_data", {}) if isinstance(p, dict) else {}
+        input_images.append(multi_modal_data.get("image"))
 
     # ... rest of generation logic
 ```
 
-For an image editing model, an example `OmniDiffusionRequest` is like:
+For an image editing model, the request `prompt` can be a dict like:
 ```python
 {
     "prompt": "turn this cat to a dog",
@@ -472,12 +475,12 @@ def get_your_model_pre_process_func(
     def pre_process_func(
         request: OmniDiffusionRequest,
         ):
-        for i, prompt in enumerate(request.prompts):
-            multi_modal_data = prompt.get("multi_modal_data", {}) if not isinstance(prompt, str) else None
-            raw_image = multi_modal_data.get("image", None) if multi_modal_data is not None else None
-            # image pre-processing
-            # after pre-processing, update the request attributes
-            ...
+        prompt = request.prompt
+        multi_modal_data = prompt.get("multi_modal_data", {}) if not isinstance(prompt, str) else None
+        raw_image = multi_modal_data.get("image", None) if multi_modal_data is not None else None
+        # image pre-processing
+        # after pre-processing, update the request attributes
+        ...
         return request
 
     return pre_process_func
@@ -923,11 +926,11 @@ When implementing a new pipeline, avoid putting all logic inside a single functi
 
 For example:
 ```
-def forward(self, req: OmniDiffusionRequest):
+def forward(self, req: DiffusionRequestBatch) -> list[DiffusionOutput]:
     prompt_embeds = self.encode_prompt(req)
     latents = self.diffuse(prompt_embeds, req)
     images = self.vae.decode(latents)
-    return DiffusionOutput(output=images)
+    return [DiffusionOutput(output=images)]
 ```
 This allows the timing utility to measure each stage (e.g., encode_prompt, diffuse, vae.decode) separately and helps identify performance bottlenecks more easily.
 

docs/contributing/model/adding_omni_model.md

@@ -475,92 +475,64 @@ thinker_hidden_states = output.multimodal_output["24"]
 
 ### Implementation Example
 
-Create stage transition processors in `vllm_omni/model_executor/stage_input_processors/your_model_name.py`:
+Create stage transition processors in `vllm_omni/model_executor/stage_input_processors/your_model_name.py`. Each inter-stage edge should provide a **coherent processor set** rather than a single monolithic function:
+
+| Suffix | Role | Runs when |
+|--------|------|-----------|
+| `*_full_payload` | Worker-side payload producer | `async_chunk=false`; accumulates tensors and ships via connector |
+| `*_async_chunk` | Scheduler-side streaming producer | `async_chunk=true`; emits per-chunk payloads |
+| `*_token_only` | Orchestrator placeholder builder | `async_chunk=false`; allocates downstream prompt slots only |
 
 ```python
-# qwen3_omni.py
+# qwen3_omni.py (Thinker → Talker, non-async path)
 
-def thinker2talker(
-    stage_list: list[Any],
-    engine_input_source: list[int],
+def thinker2talker_token_only(
+    source_outputs: list[Any],
     prompt: OmniTokensPrompt | TextPrompt | None = None,
     requires_multimodal_data: bool = False,
+    streaming_context: Any | None = None,
 ) -> list[OmniTokensPrompt]:
-    """
-    Process thinker outputs to create talker inputs.
-
-    Args:
-        stage_list: List of stage objects
-        engine_input_source: Source stage IDs (typically [0] for thinker)
-        prompt: Original prompt data
-
-    Returns:
-        List of OmniTokensPrompt for talker stage
-    """
-    source_stage_id = engine_input_source[0]
-    thinker_outputs = stage_list[source_stage_id].engine_outputs
-    talker_inputs = []
-
-    for thinker_output in thinker_outputs:
-        output = thinker_output.outputs[0]
-        # Extract thinker embeddings and hidden states
-        thinker_prefill_embeddings = output.multimodal_output["0"].float().clone().detach().cuda()
-        thinker_hidden_states = output.multimodal_output["24"].float().clone().detach().cuda()
-
-        info = {
-            "thinker_prefill_embeddings": thinker_prefill_embeddings,
-            "thinker_hidden_states": thinker_hidden_states,
-            "thinker_sequences": thinker_output.prompt_token_ids + output.token_ids,
-            "thinker_input_ids": thinker_output.prompt_token_ids,
-        }
-
-        talker_inputs.append(
-            OmniTokensPrompt(
-                prompt_token_ids=[0] * computed_length,
-                additional_information=info,
-                multi_modal_data=None,
-            )
-        )
-
-    return talker_inputs
+    """Allocate talker prefill slots; bulk tensors arrive via the connector."""
+    ...
 
 
-def talker2code2wav(
-    stage_list: list[Any],
-    engine_input_source: list[int],
-    prompt: OmniTokensPrompt | TextPrompt | None = None,
-    requires_multimodal_data: bool = False,
-) -> list[OmniTokensPrompt]:
-    """
-    Process talker outputs to create code2wav inputs.
-    """
-    source_stage_id = engine_input_source[0]
-    talker_outputs = stage_list[source_stage_id].engine_outputs
-    code2wav_inputs = []
+def thinker2talker_full_payload(
+    transfer_manager: Any,
+    pooling_output: dict[str, Any],
+    request: OmniEngineCoreRequest,
+) -> dict[str, Any] | None:
+    """Pack accumulated thinker hidden states into OmniPayload for stage-1."""
+    ...
 
-    for talker_output in talker_outputs:
-        output = talker_output.outputs[0]
-        # Extract codec codes
-        codec_codes = (
-            output.multimodal_output["code_predictor_codes"]
-            .to(torch.long)
-            .transpose(0, 1)
-            .cpu()
-            .to(torch.long)
-            .reshape(-1)
-            .tolist()
-        )
 
-        code2wav_inputs.append(
-            OmniTokensPrompt(
-                prompt_token_ids=codec_codes,
-                multi_modal_data=None,
-            )
-        )
-
-    return code2wav_inputs
+def thinker2talker_async_chunk(
+    transfer_manager: Any,
+    multimodal_output: OmniPayload | dict[str, Any],
+    request: OmniEngineCoreRequest,
+    is_finished: bool = False,
+) -> OmniPayloadStruct | None:
+    """Stream thinker rows to talker while async_chunk is enabled."""
+    ...
 ```
 
+Wire these in `pipeline.py`:
+
+```python
+StagePipelineConfig(
+    stage_id=0,
+    custom_process_next_stage_input_func=f"{_PROC}.thinker2talker_full_payload",
+    async_chunk_process_next_stage_input_func=f"{_PROC}.thinker2talker_async_chunk",
+    ...
+),
+StagePipelineConfig(
+    stage_id=1,
+    sync_process_input_func=f"{_PROC}.thinker2talker_token_only",
+    ...
+),
+```
+
+Do **not** add a no-suffix `thinker2talker` when a `sync_process_input_func` is already declared — `_select_processor_funcs()` always prefers the `*_token_only` hook in non-async mode, so the bare function would never run. See `docs/design/rfc_stage_input_processors_refactor.md` for the full contract.
+
 ## Testing
 
 For comprehensive testing guidelines, please refer to the [Test File Structure and Style Guide](../ci/tests_style.md).

docs/contributing/profiling.md

@@ -286,3 +286,54 @@ Recommended viewers:
 - Nsight Systems GUI for CUDA kernel timelines
 
 For upstream background on the underlying vLLM profiling infrastructure, see the [vLLM profiling guide](https://docs.vllm.ai/en/stable/contributing/profiling/).
+
+## 7. Orchestrator Monitor
+
+> **Warning:** Diagnostic only. Adds lightweight counters on the orchestrator poll loop and per-replica queue depth. Disable in production unless you are actively debugging orchestrator saturation.
+
+Multi-stage omni pipelines route client-facing outputs and inter-stage connector traffic through a single-process orchestrator. When that loop is saturated (for example by large multimodal payloads on the poll path), TTFT and stability can degrade even when GPU stages are healthy. The orchestrator monitor records time-series signals to confirm whether the orchestrator or downstream stage queues are the bottleneck.
+
+Enable on the stage-0 (orchestrator) process:
+
+```bash
+vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091 \
+  --stage-id 0 \
+  --enable-orch-monitor
+```
+
+Optional output path override:
+
+```bash
+export VLLM_OMNI_ORCH_MONITOR_PATH=/tmp/vllm_omni_orch_monitor.json
+```
+
+If unset, the monitor writes `./vllm_omni_orch_monitor_<MMDDHHMM>.json` at shutdown.
+
+### Recorded metrics
+
+Each 1-second window records:
+
+| Series | Description |
+|---|---|
+| `windows.duration_s` | Wall time covered by the window |
+| `windows.loop_idle` / `windows.loop_active` | Orchestrator poll-loop iterations with no work vs. active forwarding |
+| `replicas.<stage,replica>.outputs_queue_size` | MP client `outputs_queue` backlog for that replica |
+| `replicas.<stage,replica>.inflight` | Requests currently bound/routed to the replica |
+
+On shutdown the server also logs a short summary (`loop_active_pct`, per-replica queue averages/maxima).
+
+### Relationship to other diagnostics
+
+This monitor is intentionally separate from the existing profiling tools:
+
+| Tool | Scope | Output |
+|---|---|---|
+| `--enable-diffusion-pipeline-profiler` | Diffusion stage functions (`vae.decode`, `diffuse`, …) | Per-request `stage_durations` / logs |
+| `--enable-ar-profiler` | AR stage generation time | Per-request `stage_durations` |
+| `profiler_config` (`torch` / `cuda`) | GPU/CPU kernels inside a stage worker | `trace.json`, Nsight, operator tables |
+| Prometheus `/metrics` (`vllm:omni_*`) | Pipeline SLOs and cross-stage transfer | Continuous scrape endpoint |
+| `--enable-orch-monitor` | Orchestrator poll loop + replica queue backlog | Single JSON file at shutdown |
+
+The orchestrator monitor does not use `torch.profiler` because the bottleneck signal lives in the orchestrator process (poll-loop duty cycle and queue depth), not inside stage workers. It complements Prometheus metrics when you need a post-run time series for a benchmark or regression run.
+
+Process liveness monitors (`monitor_engine_liveness`, worker/engine-core monitors) are unrelated: they detect crashed workers, not orchestrator load.

docs/design/feature/cfg_parallel.md

@@ -126,16 +126,16 @@ Call `self.diffuse` in your pipeline's forward function:
 
 ```python
 import torch.nn as nn
+from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
+
 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
+        req: DiffusionRequestBatch,
+    ) -> list[DiffusionOutput]:
+        # Read prompts and sampling params from the request
+        ...
+        # Encode prompts, initialize latents, get timesteps
         ...
         # Run diffusion loop (calls the mixin's diffuse method)
         latents = self.diffuse(

docs/design/feature/diffusion_continuous_batching.md

@@ -38,9 +38,13 @@ With continuous batching enabled:
 The current implementation is conservative:
 
 - only compatible requests are batched together
-- request-mode diffusion still runs with `max_num_seqs=1`
 - per-request progress and completion remain independent
 
+Here, "continuous batching" means the step-wise path enabled by
+`step_execution=True`. Request-mode `DiffusionRequestBatch` is static
+request-level batching for one full pipeline `forward()` call; it does not
+admit or remove requests between denoise steps.
+
 ## Enablement
 
 Use `--step-execution` as the feature gate, then increase `--max-num-seqs`
@@ -72,7 +76,7 @@ which is built from shape-sensitive and CFG-sensitive sampling fields. This is
 the core correctness rule for batching: requests are only co-batched when they
 share the same denoise tensor contract.
 
-There are two important details:
+There are three important details:
 
 - `num_inference_steps` is not part of the key, so requests with different
   total step counts can still share a batch
@@ -88,8 +92,10 @@ key also covers LoRA identity (`lora_int_id`, `lora_scale`), so requests
 targeting different adapters or scales run in separate batches and the
 worker can activate exactly one adapter per step.
 
-The current batching unit is one `OmniDiffusionRequest`. Requests with
-multiple prompts do not participate in batching today.
+The scheduler batching unit is one logical `OmniDiffusionRequest`. In the
+step-wise path, runtime tensor batching is represented as `StepInputBatch`. For
+request-mode prompt semantics, see
+[Request-Level Batching](../../user_guide/diffusion/request_batching.md).
 
 ## Runner
 
@@ -126,17 +132,15 @@ request-local scheduler state and outputs.
 the background loop and async add-request path needed for multiple requests to
 accumulate in the scheduler.
 
-This is supporting infrastructure, not the main design point. The batching
-behavior is defined by scheduler-side compatibility gating and runner-side
-batch packing.
+When `step_execution=True`, the engine routes work through the step-wise
+executor path. The continuous batching behavior is defined by scheduler-side
+compatibility gating and runner-side `StepInputBatch` packing.
 
 ## Current Limitations
 
 - Experimental feature; use `max_num_seqs=1` for the older conservative path.
 - Only native pipelines that already support `step_execution=True`.
-- Request-mode diffusion still clamps `max_num_seqs` back to `1`.
 - Only homogeneous batches keyed by `SamplingParamsKey` are supported.
-- Multi-prompt requests are not batched.
 - `cache_backend`, KV transfer, and other request-mode extras are not wired
   into the batched step-wise path yet.
 - Future work can relax the current same-shape restriction with richer

docs/design/feature/diffusion_request_level_batching.md

@@ -0,0 +1,175 @@
+# Request-Level Batching for Diffusion
+
+This document describes the request-mode batching path for diffusion pipelines.
+For end-user enablement and tuning, see
+[Request-Level Batching](../../user_guide/diffusion/request_batching.md).
+
+This is separate from
+[Continuous Batching for Step-Wise Diffusion](diffusion_continuous_batching.md).
+Request-level batching runs one full pipeline `forward()` over a static batch of
+compatible requests. Step-wise continuous batching admits work between denoise
+steps when `step_execution=True`.
+
+## Why It Helps
+
+The request-level design avoids coupling several logical requests to one request
+object. This keeps request identity, abort/error handling, and per-request
+metadata unambiguous while still allowing one fused pipeline forward pass for
+bursty or concurrent traffic.
+
+## Overview
+
+With request-level batching enabled:
+
+- each `OmniDiffusionRequest` contains one `prompt` and one `request_id`
+- the scheduler groups compatible waiting requests into one scheduler wave
+- `DiffusionRequestBatch` wraps the scheduled requests for pipeline `forward()`
+- batch-capable pipelines return `list[DiffusionOutput]`, one output per
+  request
+- `BatchRunnerOutput` maps each result back to its original `request_id`
+
+Pipelines opt in with `supports_request_batch = True` and a `forward()` method
+that accepts `DiffusionRequestBatch` and returns `list[DiffusionOutput]`.
+Pipelines that do not opt in keep the existing per-request execution path.
+
+## Enablement
+
+Request-level batching is the request-mode path, so `step_execution` must remain
+disabled. Increase `max_num_seqs` above `1` to let the scheduler keep multiple
+compatible requests active:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni \
+  --port 8091 \
+  --max-num-seqs 4
+```
+
+For bursty online ingress, `request_batch_max_wait_ms` can add a bounded
+admission wait before the first `schedule()` of a scheduler wave:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni \
+  --port 8091 \
+  --max-num-seqs 4 \
+  --request-batch-max-wait-ms 20
+```
+
+`request_batch_max_wait_ms=0` disables this wait and is the default.
+
+## Request Contract
+
+`OmniDiffusionRequest` represents one logical request. It owns one prompt,
+sampling parameters, request id, and request-local metadata. Runtime batches are
+formed by the scheduler and represented separately from the request payload.
+
+Runtime batching is represented by:
+
+- [`DiffusionSchedulerOutput`](gh-file:vllm_omni/diffusion/sched/interface.py)
+  for scheduled request ids and request payloads
+- [`DiffusionRequestBatch`](gh-file:vllm_omni/diffusion/worker/request_batch.py)
+  for the pipeline-facing request batch
+- [`BatchRunnerOutput`](gh-file:vllm_omni/diffusion/worker/utils.py) for
+  per-request results
+
+`DiffusionRequestBatch` intentionally exposes compatibility properties such as
+`prompts`, `sampling_params`, `request_id`, and `kv_sender_info` so migrated
+pipelines can stay close to upstream code while using a batch-aware contract.
+
+## Scheduler
+
+The scheduler derives its capacity from `max_num_seqs` through
+`max_num_running_reqs`. It exposes waiting/running queue counters so the engine
+can decide whether admission wait is useful before scheduling a new wave.
+
+Batch compatibility is controlled by
+[`SamplingParamsKey`](gh-file:vllm_omni/diffusion/sched/interface.py). The key
+contains shape-sensitive and guidance-sensitive fields, including output count
+and LoRA identity. Requests with incompatible shapes, CFG settings, output
+counts, LoRA adapters, or LoRA scales are kept in separate batches.
+
+Admission is conservative:
+
+- the scheduler only batches compatible requests
+- FIFO ordering is preserved
+- an incompatible request at the head of the waiting queue blocks later
+  compatible requests
+
+## Engine
+
+[`DiffusionEngine`](gh-file:vllm_omni/diffusion/diffusion_engine.py) resolves
+request-batch capability during initialization from the configured pipeline
+class, including custom pipeline classes.
+
+The capability check uses the pipeline class attribute
+`supports_request_batch = True`. Pipelines that set this attribute must implement
+a request-batch-compatible `forward()` contract and return one
+`DiffusionOutput` per request; the runner validates that return shape at runtime.
+
+When the selected pipeline is batch-capable and `step_execution=False`, request
+mode routes scheduler waves through the batch executor path. Otherwise it keeps
+the per-request executor path.
+
+The optional admission wait runs only when:
+
+- request batching is supported
+- `step_execution=False`
+- `request_batch_max_wait_ms > 0`
+- no requests are currently running
+
+The wait exits early when the waiting queue reaches capacity, when the queue is
+stable for a short window, when the deadline expires, or when the engine stops.
+
+## Executor And Runner
+
+The executor exposes two request-mode entries:
+
+- `execute_request`: one worker call per scheduled request
+- `execute_batch`: one worker call for the whole `DiffusionSchedulerOutput`
+
+On the batch path, the worker builds a `DiffusionRequestBatch` and runs the
+pipeline once. Request-local setup remains per request:
+
+- KV transfer metadata
+- random generator and seed handling
+- request output/error/abort mapping
+
+Shared batch setup happens once per batch when possible:
+
+- cache refresh
+- LoRA activation for the homogeneous adapter key
+- pipeline `forward(req_batch)`
+
+Large tensor IPC still uses the shared-memory packing path. The packer traverses
+both normal `RunnerOutput.result` wrappers and nested batch results so batched
+outputs do not fall back to pickle IPC for tensor payloads.
+
+## Current Limitations
+
+- Only pipelines that declare the request-batch contract use fused batch
+  execution.
+- Batches are homogeneous under `SamplingParamsKey`; heterogeneous resolution or
+  incompatible guidance settings do not co-batch yet.
+- FIFO scheduling can reduce batching opportunities when an incompatible
+  request is at the front of the queue.
+- `request_batch_max_wait_ms` improves burst coalescing but can add latency to
+  the first request in a scheduler wave. Keep it small for latency-sensitive
+  serving.
+- Step-wise continuous batching is documented separately and only applies when
+  `step_execution=True`.
+
+## Related Files
+
+- Request object and request batch:
+  [`vllm_omni/diffusion/request.py`](gh-file:vllm_omni/diffusion/request.py)
+- Scheduler interface:
+  [`vllm_omni/diffusion/sched/interface.py`](gh-file:vllm_omni/diffusion/sched/interface.py)
+- Scheduler base:
+  [`vllm_omni/diffusion/sched/base_scheduler.py`](gh-file:vllm_omni/diffusion/sched/base_scheduler.py)
+- Engine:
+  [`vllm_omni/diffusion/diffusion_engine.py`](gh-file:vllm_omni/diffusion/diffusion_engine.py)
+- Worker runner:
+  [`vllm_omni/diffusion/worker/diffusion_model_runner.py`](gh-file:vllm_omni/diffusion/worker/diffusion_model_runner.py)
+- Executor interface:
+  [`vllm_omni/diffusion/executor/abstract.py`](gh-file:vllm_omni/diffusion/executor/abstract.py)
+- Tests:
+  [`tests/diffusion/test_diffusion_engine.py`](gh-file:tests/diffusion/test_diffusion_engine.py)

docs/design/feature/vae_parallel.md

@@ -448,6 +448,42 @@ Complete examples in the codebase:
 
 ---
 
+## Spatially-Sharded Decode (Wan)
+
+The tile-parallel executor above assigns independent spatial **tiles** to ranks. The Wan VAE additionally supports a **spatially-sharded decode** backend, selected through `DiffusionParallelConfig.vae_parallel_mode` (`"spatial_shard_height"` or `"spatial_shard_width"`; default is `"tile"`).
+
+### How it differs from tile parallel
+
+| Aspect | Tile parallel (`"tile"`) | Spatially-sharded (`"spatial_shard_height"`/`"spatial_shard_width"`) |
+|--------|--------------------------|-----------------------------------------------|
+| Unit of work | Independent overlapping tiles | A single global feature map sharded along H or W |
+| Cross-rank communication | Gather tiles to rank 0, stitch + blend | Per-conv **halo exchange** of boundary rows/cols (P2P) |
+| Output assembly | Blend overlapping tiles | All-gather shards on rank 0, trim padding (matches `broadcast_result=False`) |
+| Scope | Decode + encode | Decode only |
+
+Spatial-shard decode swaps the decoder's spatial convolutions/padding for halo-exchanging variants (`WanDistConv2d`, `WanDistCausalConv3d`, `WanDistZeroPad2d`) so each rank only holds a shard of the activations but still sees the correct receptive field at shard boundaries. Implementation lives in `vllm_omni/diffusion/distributed/autoencoders/wan_spatial_shard.py`.
+
+### Wiring
+
+`vae_parallel_mode` flows through the same path as `vae_patch_parallel_size`:
+
+```text
+serve.py (--vae-parallel-mode) / OmniEngineArgs
+  -> DiffusionParallelConfig.vae_parallel_mode
+  -> registry.py: model.vae.set_parallel_size(vae_pp_size, mode=...)
+  -> DistributedVaeExecutor.parallel_mode
+  -> DistributedAutoencoderKLWan.tiled_decode dispatch
+```
+
+`DistributedAutoencoderKLWan._spatial_shard_decode_enabled()` gates the path: it requires distributed decode to be enabled, a 5D latent, and `vae_patch_parallel_size == DiT group size`. Otherwise it logs a warning and falls back to tile-parallel decode.
+
+### Notes
+
+- The decoder is patched **in place** the first time spatial-shard decode runs and is bound to a single split dimension for the lifetime of the VAE instance.
+- Numerical correctness vs. single-GPU decode is covered by `tests/diffusion/distributed/test_wan_spatial_shard.py::test_spatial_shard_decode_matches_reference` (multi-GPU, nightly `full_model` + `distributed_cuda`).
+
+---
+
 ## Summary
 
 Adding VAE Patch Parallel support to diffusion model:

docs/design/index.md

@@ -11,6 +11,7 @@ This section contains design documents and architecture specifications for vLLM-
 - [Disaggregated Inference](feature/disaggregated_inference.md)
 - [Ray-based Execution](feature/ray_based_execution.md)
 - [Adding Step Execution Support for Diffusion Pipelines](feature/diffusion_step_execution.md)
+- [Request-Level Batching for Diffusion](feature/diffusion_request_level_batching.md)
 - [Continuous Batching for Step-Wise Diffusion](feature/diffusion_continuous_batching.md)
 
 ## Infrastructure Design Documents

docs/design/module/dit_module.md

@@ -199,7 +199,7 @@ class _BaseScheduler(SchedulerInterface):
 - **Shared cleanup logic**: Duplicate-request checks, finish handling, and state removal are centralized instead of
   duplicated in each policy.
 
-- **Current constraint boundary**: `_BaseScheduler` derives `max_num_running_reqs` from `max_num_seqs`, but request-mode diffusion is still clamped back to `1` by the engine. The step-wise path can keep this above `1` for compatible-request batching.
+- **Current constraint boundary**: `_BaseScheduler` derives `max_num_running_reqs` from `max_num_seqs`. Request-mode diffusion can use that capacity for compatible independent requests when the configured pipeline declares `supports_request_batch = True`; step-wise diffusion uses the same scheduler capacity for compatible step batches.
 
 #### 2.4 Current `RequestScheduler` Policy
 
@@ -215,7 +215,7 @@ class RequestScheduler(_BaseScheduler):
 
 - **FIFO request scheduling**: Waiting requests are promoted in queue order.
 
-- **Single-request admission**: `RequestScheduler` still admits one active request at a time because request-mode execution completes a whole request per dispatch.
+- **Compatible request admission**: `RequestScheduler` admits waiting requests while capacity remains and the request's `SamplingParamsKey` is compatible with the active batch. Request-mode execution keeps each logical request independent, while batch-capable pipelines receive the scheduled requests as a runner-side `DiffusionRequestBatch`.
 
 - **Executor result feedback**: `update_from_output()` converts executor output into `FINISHED_COMPLETED` or `FINISHED_ERROR` and returns finished request ids.
 
@@ -235,7 +235,12 @@ while True:
 
 - **No scheduler-owned IPC**: Scheduler no longer talks to workers directly.
 
-- **Split concurrency model**: Request-mode diffusion remains single-active-request, while the step-wise path can keep multiple compatible requests running and advance them independently between denoise steps.
+- **Split concurrency model**: Request-mode diffusion can schedule a static
+  batch of compatible independent requests for one full pipeline forward when
+  the pipeline supports request-level batching; the step-wise path can also
+  admit or remove compatible requests between denoise steps. See
+  [Request-Level Batching](../feature/diffusion_request_level_batching.md) and
+  [Continuous Batching for Step-Wise Diffusion](../feature/diffusion_continuous_batching.md).
 
 ---
 

docs/features/custom_pipeline.md

@@ -54,20 +54,24 @@ Create a custom pipeline class that extends an existing pipeline. In this exampl
 
 ```python
 # custom_pipeline.py
-from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
+from vllm_omni.diffusion.data import OmniDiffusionConfig
 from vllm_omni.diffusion.models.qwen_image.pipeline_qwen_image_edit import QwenImageEditPipeline
+from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
 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):
+    def forward(self, req: DiffusionRequestBatch):
+        # Optionally customize sampling parameters
+        actual_num_steps = req.sampling_params.num_inference_steps or 50
+        req.sampling_params.num_inference_steps = actual_num_steps
+
         # Call parent's forward to get normal output
-        output = super().forward(req=req, prompt=prompt, negative_prompt=negative_prompt, **kwargs)
+        output = super().forward(req=req)
 
         # 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)
 

docs/getting_started/installation/gpu.md

@@ -34,7 +34,7 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr
 
 ### Pre-built wheels
 
-Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, and 0.23.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
+Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, 0.23.0, and 0.24.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
 
 === "NVIDIA CUDA"
 

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

@@ -20,7 +20,7 @@ Therefore, it is recommended to install vLLM and vLLM-Omni with a **fresh new**
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.23.0 --torch-backend=auto
+uv pip install vllm==0.24.0 --torch-backend=auto
 ```
 
 #### Installation of vLLM-Omni
@@ -39,13 +39,13 @@ uv pip install 'vllm-omni[demo]'
 # --8<-- [start:build-wheel-from-source]
 
 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.23.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.24.0 release version of the library
 
 ```bash
-uv pip install vllm==0.23.0 --torch-backend=auto
+uv pip install vllm==0.24.0 --torch-backend=auto
 ```
 
-The 0.23.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The 0.24.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -66,12 +66,12 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.23.0
+git checkout v0.24.0
 ```
 Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
 ```bash
-#For CUDA 13.0 (the default for v0.23.0; the wheel filename has no `+cu130` suffix)
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.23.0/vllm-0.23.0-cp38-abi3-manylinux_2_28_x86_64.whl
+#For CUDA 13.0 (the default for v0.24.0; the wheel filename has no `+cu130` suffix)
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.24.0/vllm-0.24.0-cp38-abi3-manylinux_2_28_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash
@@ -102,7 +102,7 @@ docker run --runtime nvidia --gpus 2 \
     --env "HF_TOKEN=$HF_TOKEN" \
     -p 8091:8091 \
     --ipc=host \
-    vllm/vllm-omni:v0.23.0 \
+    vllm/vllm-omni:v0.24.0 \
     vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 

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

@@ -13,7 +13,7 @@ vLLM-Omni current recommends the steps in under setup through Docker Images.
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.23.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.23.0/rocm722
+uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
 ```
 
 #### Installation of vLLM-Omni
@@ -34,13 +34,13 @@ uv pip install onnxruntime-rocm
 # --8<-- [start:build-wheel-from-source]
 
 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.23.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.24.0 release version of the library
 
 ```bash
-uv pip install vllm==0.23.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.23.0/rocm722
+uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
 ```
 
-The pre-built 0.23.0 vLLM wheel targets ROCm 7.2.2. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The pre-built 0.24.0 vLLM wheel targets ROCm 7.2.2. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -58,7 +58,7 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.23.0
+git checkout v0.24.0
 python3 -m pip install -r requirements/rocm.txt
 python3 setup.py develop
 ```
@@ -130,7 +130,7 @@ docker run --rm \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HF_TOKEN=$HF_TOKEN" \
   -p 8091:8091 \
-  vllm/vllm-omni-rocm:v0.23.0 \
+  vllm/vllm-omni-rocm:v0.24.0 \
   --model Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 
@@ -149,7 +149,7 @@ docker run --rm -it \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HF_TOKEN=$HF_TOKEN" \
   --entrypoint bash \
-  vllm/vllm-omni-rocm:v0.23.0
+  vllm/vllm-omni-rocm:v0.24.0
 ```
 
 # --8<-- [end:pre-built-images]

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

@@ -1,19 +1,16 @@
 # --8<-- [start:requirements]
 
-For detailed hardware and software requirements, please refer to the [vllm-ascend installation documentation](https://docs.vllm.ai/projects/ascend/en/latest/installation.html).
+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-release]
 
-The recommended way to use vLLM-Omni on NPU is through the vllm-ascend pre-built Docker images:
+The recommended way to use vLLM-Omni on NPU is through the vLLM-Ascend pre-built Docker images:
 
 ```bash
-# Update the vllm-ascend image
-# Atlas A2:
-# export IMAGE=quay.io/ascend/vllm-ascend:v0.18.0rc1
-# Atlas A3:
-# export IMAGE=quay.io/ascend/vllm-ascend:v0.18.0rc1-a3
-export IMAGE=quay.io/ascend/vllm-ascend:v0.18.0rc1
+# vLLM-Ascend image
+# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.24.0
+# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.24.0-a3
 docker run --rm \
     --name vllm-omni-npu \
     --shm-size=1g \
@@ -31,11 +28,11 @@ docker run --rm \
     -v /etc/ascend_install.info:/etc/ascend_install.info \
     -v /root/.cache:/root/.cache \
     -p 8000:8000 \
-    -it $IMAGE bash
+    -it quay.io/atlas-ci/vllm-ascend:0.24.0-a3 bash
 
 # Inside the container, install vLLM-Omni from source
 cd /vllm-workspace
-git clone -b v0.18.0 https://github.com/vllm-project/vllm-omni.git
+git clone -b v0.24.0 https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
 pip install -v -e . --no-build-isolation
 # or VLLM_OMNI_TARGET_DEVICE=npu pip install -v -e .
@@ -87,14 +84,11 @@ Supported images as following.
 Here's an example deployment command that has been verified on 4 x NPUs:
 
 ```bash
-# Atlas A2:
-# export IMAGE=quay.io/ascend/vllm-omni:v0.18.0
-# Atlas A3:
-# export IMAGE=quay.io/ascend/vllm-omni:v0.18.0-a3
-export IMAGE=quay.io/ascend/vllm-omni:v0.18.0
+# Atlas A2: quay.io/ascend/vllm-omni:v0.24.0
+# Atlas A3: quay.io/ascend/vllm-omni:v0.24.0-a3
 docker run --rm \
-    --name vllm-omni-npu \
-    --shm-size=1g \
+    --name vllm-omni-a3 \
+    --shm-size=4g \
     --device /dev/davinci0 \
     --device /dev/davinci1 \
     --device /dev/davinci2 \
@@ -106,10 +100,17 @@ docker run --rm \
     -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
     -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
     -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
+    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
     -v /etc/ascend_install.info:/etc/ascend_install.info \
-    -v /root/.cache:/root/.cache \
-    -p 8000:8000 \
-    -it $IMAGE bash
+    -v ~/.cache:/root/.cache \
+    -p 8091:8091 \
+    -it quay.io/ascend/vllm-omni:v0.24.0-a3 bash
 ```
 
 !!! tip

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==0.23.0 --torch-backend=auto
+uv pip install vllm==0.24.0 --torch-backend=auto
 
 # On ROCm
-uv pip install vllm==0.23.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.23.0/rocm722
+uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
 
 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
@@ -31,11 +31,10 @@ uv pip install -e .
 
 For additional installation methods — please see the [installation guide](installation/README.md).
 
-
 !!! note
     It is important to install the same major & minor version of vLLM and vLLM Omni, otherwise things may not work as expected. If the versions are misaligned, you will see a warning when you import vLLM Omni.
 
-    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.23.0` and vLLM Omni `0.23.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
+    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.24.0` and vLLM Omni `0.24.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
 
 ## Offline Inference
 
@@ -52,14 +51,14 @@ if __name__ == "__main__":
     images[0].save("coffee.png")
 ```
 
-You can pass a list of prompts and wait for them to process altogether, shown below.
+You can pass a list of prompts and wait for the independent requests to finish,
+as shown below.
 
 !!! info
 
-    However, it is not currently recommended to do so
-    because not all models support batch inference,
-    and batch requesting mostly does not provide significant performance improvement (despite the impression that it does).
-    This feature is primarily for the sake of interface compatibility with vLLM and to allow for future improvements.
+    For diffusion pipelines, each prompt becomes a separate logical request.
+    The runtime may automatically batch compatible in-flight requests through
+    the scheduler and runner.
 
 ```python
 from vllm_omni.entrypoints.omni import Omni
@@ -88,9 +87,9 @@ if __name__ == "__main__":
 
 !!! info
 
-    For diffusion pipelines, the stage config field `stage_args.[].engine_args.max_num_seqs` is 1 by default, and the input
-    list is sliced into single-item requests before feeding into the diffusion pipeline. For models that do internally support
-    batched inputs, you can [modify this configuration](../configuration/stage_configs.md) to let the model accept a longer batch of prompts.
+    For diffusion request-level batching controls such as `max_num_seqs` and
+    `request_batch_max_wait_ms`, see
+    [Request-Level Batching](../user_guide/diffusion/request_batching.md).
 
 For more usages, please refer to [offline inference](../user_guide/examples/offline_inference/qwen2_5_omni.md)
 

docs/serving/speech_api.md

@@ -122,7 +122,8 @@ Content-Type: application/json
 | `max_new_tokens` | integer | 2048 | Maximum tokens to generate |
 | `initial_codec_chunk_frames` | integer | null | Per-request initial chunk size override for TTFA tuning. When null, IC is computed dynamically based on server load. |
 | `non_streaming_mode` | bool | null | Qwen3-TTS prompt construction mode override. Does not affect HTTP response streaming or async-chunk pipelining. When null, Qwen3-TTS uses model defaults: Base=false, CustomVoice/VoiceDesign=true. |
-| `stream` | bool | false | Stream raw PCM chunks as they are decoded (requires `response_format="pcm"`) |
+| `stream` | bool | false | When true, stream OpenAI `speech.audio.*` SSE events (requires `response_format="pcm"` or `"wav"`). For raw PCM/WAV byte streaming, set `stream_format="audio"`. |
+| `stream_format` | string | null | Streaming output format. `"audio"` streams raw audio bytes as they are decoded; `"sse"` streams OpenAI `speech.audio.*` Server-Sent Events. If omitted, `stream=true` selects SSE and `stream=false` remains non-streaming. See [Response Format](#response-format). |
 
 **Supported languages:** Only applicable to Qwen3-TTS. Derived from the model configuration (`talker_config.codec_language_id` in the checkpoint's `config.json`), plus `Auto`, which is always accepted. Official Qwen3-TTS checkpoints support: Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.
 
@@ -136,7 +137,57 @@ Content-Type: application/json
 
 ### Response Format
 
-Returns binary audio data with appropriate `Content-Type` header (e.g., `audio/wav`).
+The response shape depends on the streaming parameters:
+
+**Non-streaming (default).** With `stream=false` and no `stream_format`, returns the
+complete clip as binary audio data with an appropriate `Content-Type` header (e.g.
+`audio/wav`). The raw-bytes body has no JSON carrier, so no `usage` is reported.
+
+**Raw audio stream** (`stream_format="audio"`). Streams raw audio bytes (PCM or
+WAV) as they are decoded.
+
+Both streaming modes (`stream_format="audio"` and `"sse"`) require
+`response_format="pcm"` or `"wav"`, and `speed` must be `1.0` (or omitted).
+
+**SSE stream** (`stream=true` or `stream_format="sse"`). Streams [OpenAI
+`speech.audio.*` Server-Sent Events](https://platform.openai.com/docs/api-reference/audio-streaming).
+Each event has an `event:` line and a JSON `data:` line:
+
+- `speech.audio.delta` — a base64 audio chunk:
+
+    ```json
+    { "type": "speech.audio.delta", "audio": "<base64>", "response_format": "pcm" }
+    ```
+
+- `speech.audio.done` — terminal event, carrying token `usage`:
+
+    ```json
+    {
+        "type": "speech.audio.done",
+        "usage": {
+            "input_tokens": 119,
+            "output_tokens": 77,
+            "total_tokens": 196,
+            "input_token_details": { "text_tokens": 18, "audio_tokens": 101 }
+        }
+    }
+    ```
+
+- `speech.audio.error` — emitted instead of `speech.audio.done` if generation fails:
+
+    ```json
+    { "type": "speech.audio.error", "error": { "message": "...", "type": "server_error", "param": null, "code": 500 } }
+    ```
+
+The `usage` object on `speech.audio.done` is the same shape returned per item by the
+[batch endpoint](#batch-speech-generation):
+
+- `input_tokens` = `text_tokens` + `audio_tokens`
+    - `text_tokens`: tokens of the synthesized text (`input` plus `instructions`)
+    - `audio_tokens`: reference-audio codec frames, non-zero only for in-context
+      voice cloning (Base task); `0` for CustomVoice/VoiceDesign or x-vector-only
+- `output_tokens`: generated codec tokens
+- `total_tokens` = `input_tokens` + `output_tokens`
 
 ### Voices Endpoint
 
@@ -490,7 +541,13 @@ Each item in the `items` array requires only `input` (the text). All other field
             "index": 0,
             "status": "success",
             "audio_data": "<base64-encoded audio>",
-            "media_type": "audio/wav"
+            "media_type": "audio/wav",
+            "usage": {
+                "input_tokens": 119,
+                "output_tokens": 77,
+                "total_tokens": 196,
+                "input_token_details": { "text_tokens": 18, "audio_tokens": 101 }
+            }
         },
         {
             "index": 1,
@@ -504,6 +561,18 @@ Each item in the `items` array requires only `input` (the text). All other field
 }
 ```
 
+Each successful item carries a `usage` object (errored items omit it):
+
+- `input_tokens` = `text_tokens` + `audio_tokens`
+    - `text_tokens`: tokens of the synthesized text (`input` plus `instructions`)
+    - `audio_tokens`: reference-audio codec frames, non-zero only for in-context
+      voice cloning (Base task); `0` for CustomVoice/VoiceDesign or x-vector-only
+- `output_tokens`: generated codec tokens
+- `total_tokens` = `input_tokens` + `output_tokens`
+
+This is the same `usage` object emitted on the terminal `speech.audio.done` event
+of the single endpoint's [SSE stream](#response-format) (`stream_format="sse"`).
+
 ### Examples
 
 **Basic batch with shared defaults:**

docs/user_guide/diffusion/attention_backends.md

@@ -20,6 +20,8 @@ The full set of backends and their platform defaults is in the **Backend Options
 | `TORCH_SDPA` | PyTorch `scaled_dot_product_attention` with the default backend dispatcher. Most conservative; always available. |
 | `SAGE_ATTN` | SageAttention 2.2 — INT8-quantized attention with FP16 accumulation. Lossy but typically visually indistinguishable on diffusion outputs. Requires `sageattention`. |
 | `SAGE_ATTN_3` | Requires `sageattn3` from `SageAttention/sageattention3_blackwell`. CUDA only, intended for Blackwell GPUs, with GQA/MQA requests falling back to PyTorch SDPA. |
+| `FLASH_ATTN_HUB` | FlashAttention 2 from HuggingFace `kernels` library. Useful for train/rollout alignment. |
+| `FLASH_ATTN_3_HUB` | FlashAttention 3 from HuggingFace `kernels` library. CUDA Hopper (sm_90+) only; falls back to `FLASH_ATTN_HUB` on older GPUs. |
 
 
 ## Configuration
@@ -210,6 +212,36 @@ Notes:
 - `SAGE_ATTN_3` is only selected on CUDA when `sageattn3` is importable and the GPU is Blackwell-class.
 - SageAttention3's Blackwell kernel assumes `Hq == Hkv`. In vLLM-Omni, GQA/MQA diffusion requests fall back to PyTorch SDPA for correctness.
 
+## HuggingFace Kernels Hub Backends
+
+To achieve perfect numerical consistency between **training** (typically using HuggingFace Diffusers with Hub kernels) and **serving/rollout** (using vLLM-Omni), you can use the Hub-based attention backends. This eliminates numerical drift / sampling divergence caused by executing different local kernel versions during rollout.
+
+The following backend options are supported:
+- `FLASH_ATTN_HUB` (HuggingFace `kernels-community/flash-attn2`)
+- `FLASH_ATTN_3_HUB` (HuggingFace `kernels-community/flash-attn3`, Hopper sm_90+ only)
+
+### Installation
+
+To use these backends, you must install the `kernels` library:
+
+```bash
+pip install kernels==0.14.1
+```
+
+If the `kernels` library is not available in the environment, vLLM-Omni will log a warning and fall back gracefully to the corresponding local backend implementations (`FLASH_ATTN`). On CUDA GPUs below Hopper (compute capability < 9.0), `FLASH_ATTN_3_HUB` falls back to `FLASH_ATTN_HUB`.
+
+### Usage
+
+Select a Hub backend using the global CLI flag or environment variables:
+
+```bash
+# Environment variable
+export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN_HUB
+
+# CLI flag
+vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN_HUB
+```
+
 ## Usage Examples
 
 ### Default (auto-route)

docs/user_guide/diffusion/parallelism/overview.md

@@ -10,7 +10,7 @@ This guide covers the parallelism methods in vLLM-Omni for speeding up diffusion
 | **[Sequence Parallelism](sequence_parallel.md)**   | Splits sequence dimension across GPUs (Ulysses-SP, Ring-Attention, or hybrid) for high-resolution images and videos |
 | **[CFG-Parallel](cfg_parallel.md)**                | Runs CFG positive/negative branches on separate GPUs for ~1.8x speedup on guided generation                         |
 | **[Pipeline Parallelism](pipeline_parallel.md)**   | Splits the denoising transformer block-wise across sequential GPU stages to reduce per-GPU model memory             |
-| **[VAE Patch Parallelism](vae_patch_parallel.md)** | Distributes VAE decode spatially across GPUs to reduce peak VAE memory                                              |
+| **[VAE Parallelism](vae_parallelism.md)** | Distributes VAE decode spatially across GPUs to reduce peak VAE memory                                              |
 | **[HSDP](hsdp.md)**                                | Shards full model weights via PyTorch FSDP2 to enable large-model inference on memory-constrained GPUs              |
 | **[Expert Parallelism](expert_parallel.md)**       | Shards MoE expert blocks across GPUs for MoE models (e.g. HunyuanImage3.0)                                          |
 

docs/user_guide/diffusion/parallelism/vae_parallelism.md

@@ -1,4 +1,4 @@
-# VAE Patch Parallelism Guide
+# VAE Parallelism Guide
 
 
 ## Table of Content
@@ -15,7 +15,7 @@
 
 ## Overview
 
-VAE Patch Parallelism distributes the VAE (Variational AutoEncoder) decode/encode computation across multiple GPUs by splitting the latent space into spatial tiles or patches. Each GPU processes a subset of tiles in parallel, significantly reducing peak memory consumption during the VAE decode stage while maintaining output quality.
+VAE parallelism distributes VAE (Variational AutoEncoder) decode/encode work across multiple GPUs. This guide covers VAE patch/tile parallelism, which splits latent space into spatial tiles or patches, and Wan spatial-shard decode, which shards decoder feature maps along height or width.
 
 This is particularly useful for:
 - **High-resolution image generation** where VAE decode can become a memory bottleneck
@@ -25,7 +25,7 @@ This is particularly useful for:
 See supported models list in [Supported Models](../../diffusion_features.md#supported-models).
 
 
-VAE Patch Parallelism uses two strategies based on image size:
+VAE patch parallelism uses two strategies based on image size:
 
 | Strategy | Use Case | How It Works | Overlap Handling | Output Quality |
 |----------|----------|--------------|------------------|----------------|
@@ -113,6 +113,7 @@ In `DiffusionParallelConfig`:
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
 | `vae_patch_parallel_size` | int | 1 | Number of GPUs for VAE patch/tile parallelism. Set to 2 or higher to enable. Should typically match `tensor_parallel_size` as they share the same process group. |
+| `vae_parallel_mode` | str | `"tile"` | VAE parallel decode strategy: `"tile"` (default tile/patch parallel decode), `"spatial_shard_height"`, or `"spatial_shard_width"` (spatially-sharded decode, Wan only). See [Spatially-Sharded Decode](#spatially-sharded-decode-wan). |
 
 Additional requirements:
 
@@ -125,6 +126,51 @@ Additional requirements:
 
 ---
 
+## Spatially-Sharded Decode (Wan)
+
+The default `vae_parallel_mode="tile"` distributes whole tiles across ranks. For the **Wan** VAE there is an alternative decode strategy, **spatially-sharded decode**, selected via `vae_parallel_mode="spatial_shard_height"` or `vae_parallel_mode="spatial_shard_width"`.
+
+Instead of assigning independent tiles to ranks, spatial-shard decode shards the decoder feature maps along the height (`spatial_shard_height`) or width (`spatial_shard_width`) dimension and exchanges halo rows/columns between neighboring ranks around the spatial convolutions. This keeps the receptive field correct across shard boundaries, so the result matches the single-GPU decode within numerical tolerance.
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+    parallel_config=DiffusionParallelConfig(
+        tensor_parallel_size=2,
+        vae_patch_parallel_size=2,               # must match the DiT group size
+        vae_parallel_mode="spatial_shard_width", # or "spatial_shard_height"
+    ),
+)
+```
+
+Or from the CLI / serving entrypoint:
+
+```bash
+vllm serve Wan-AI/Wan2.1-T2V-1.3B-Diffusers --omni \
+    --tensor-parallel-size 2 \
+    --vae-patch-parallel-size 2 \
+    --vae-parallel-mode spatial_shard_width
+```
+
+**Constraints and behavior:**
+
+- Spatial-shard decode is **decode-only** and currently implemented for the **Wan** VAE. Other models ignore `spatial_shard_*` modes.
+- It requires `vae_patch_parallel_size` to **match the DiT process group size**. If it does not, the VAE logs a warning and **falls back to tile-parallel decode** at runtime.
+- `spatial_shard_height` and `spatial_shard_width` are mutually exclusive for a given VAE instance (the decoder is patched in place for a single split dimension).
+
+For end-to-end latency/throughput, launch serving with the desired `vae_parallel_mode` and use the existing diffusion serving benchmark:
+
+```bash
+python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
+    --endpoint /v1/videos --dataset random --task t2v --num-prompts 1 \
+    --height 480 --width 832 --num-frames 17 --max-concurrency 1
+```
+
+---
+
 ## Best Practices
 
 ### When to Use

docs/user_guide/diffusion/request_batching.md

@@ -0,0 +1,110 @@
+# Request-Level Batching
+
+Request-level batching lets diffusion serving combine multiple compatible
+logical requests into one pipeline `forward()` call. Each prompt remains a
+separate request with its own `request_id`, sampling parameters, seed, output,
+error, and abort state. The scheduler decides which requests can run together.
+
+!!! warning "Prompt List Semantics"
+
+    Diffusion request-level batching does not support a top-level packed
+    list-prompt request. Submit multiple prompts as independent requests and let
+    the scheduler batch compatible in-flight requests. Multimodal payloads stay
+    inside a single prompt dict, for example
+    `{"prompt": "...", "multi_modal_data": {"image": image}}`.
+
+## Enablement
+
+Increase `max_num_seqs` above `1` to allow the request scheduler to keep more
+than one compatible request active:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni \
+  --port 8091 \
+  --max-num-seqs 4
+```
+
+For bursty online traffic, you can also set a small admission wait window. This
+lets the engine wait briefly before the first `schedule()` of a new wave so
+nearby compatible requests can arrive and share the same fused forward pass:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni \
+  --port 8091 \
+  --max-num-seqs 4 \
+  --request-batch-max-wait-ms 20
+```
+
+`--request-batch-max-wait-ms 0` is the default and disables admission waiting,
+so there is no added wait latency.
+
+For deploy YAMLs, configure the diffusion stage engine args:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    stage_type: diffusion
+    engine_args:
+      max_num_seqs: 4
+      request_batch_max_wait_ms: 20
+```
+
+## Compatibility
+
+Only pipelines that declare request-batch support use the fused request-batch
+path. The engine validates that the pipeline `forward()` uses the request-batch
+contract and returns `list[DiffusionOutput]`. Pipelines that do not support this
+contract do not use fused `pipeline.forward(batch)`; scheduled requests are
+executed through per-request worker calls.
+
+The scheduler batches only compatible requests. Compatibility is based on
+shape-sensitive and guidance-sensitive sampling fields, including resolution,
+frame count, CFG settings, output count, and LoRA identity. Requests with
+different LoRA adapters or scales are kept in separate batches so the worker
+activates one adapter per batch.
+
+Request-level batching applies when `step_execution=False`. For the separate
+step-wise runtime, see [Step Execution](step_execution.md).
+
+## Tuning
+
+- `max_num_seqs` caps the number of active compatible requests in one scheduler
+  wave.
+- `request_batch_max_wait_ms` is an upper bound on extra admission wait before a
+  new wave starts. Keep it small for latency-sensitive serving; values such as
+  `10` to `50` ms are a practical starting range for bursty HTTP ingress.
+- `0` disables admission waiting and preserves the lowest first-request latency.
+- FIFO ordering is conservative: an incompatible request at the front of the
+  waiting queue can block later compatible requests from joining the current
+  batch.
+
+## Python API
+
+When constructing `Omni`, pass the same engine arguments:
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    max_num_seqs=4,
+    request_batch_max_wait_ms=20.0,
+)
+
+outputs = omni.generate(
+    [
+        "a cup of coffee on a table",
+        "a toy dinosaur on a sandy beach",
+        "a fox waking up in bed and yawning",
+    ]
+)
+```
+
+`Omni.generate([...])` submits each list item as its own logical diffusion
+request. The runtime may batch those requests internally when their sampling
+parameters are compatible.
+
+## For Contributors
+
+For implementation details and model-author guidance, see
+[Request-Level Batching for Diffusion](../../design/feature/diffusion_request_level_batching.md).

docs/user_guide/diffusion_features.md

@@ -61,7 +61,7 @@ Memory optimization methods help reduce GPU memory usage, enabling inference on
 |--------|-------------|----------|
 | **[CPU Offload](diffusion/cpu_offload_diffusion.md)** | Offloads model components to CPU memory | Limited VRAM, large models on consumer GPUs |
 | **[Quantization](quantization/overview.md)** | Reduces transformer stages from BF16 to FP8/INT8/etc. | Limited VRAM, minimal accuracy loss    |
-| **[VAE Patch Parallelism](diffusion/parallelism/vae_patch_parallel.md)** | Distributes VAE decode tiling across GPUs | High-resolution generation with reduced VAE memory peak |
+| **[VAE Parallelism](diffusion/parallelism/vae_parallelism.md)** | Distributes VAE decode work across GPUs | High-resolution generation with reduced VAE memory peak |
 
 ### Extensions
 
@@ -75,13 +75,17 @@ Extension methods add specialized capabilities to diffusion models beyond standa
 
 ### Execution Modes
 
-Execution modes control how the diffusion pipeline processes denoise steps.
+Execution modes control how the diffusion pipeline processes requests and
+denoise steps.
 
 | Method | Description | Best For |
 |--------|-------------|----------|
+| **[Request-Level Batching](diffusion/request_batching.md)** | Scheduler batches compatible independent diffusion requests into one pipeline forward pass | Bursty online serving and multi-request throughput |
 | **[Step Execution](diffusion/step_execution.md)** | Per-step denoise execution with mid-request abort support | Request cancellation between denoise steps, fine-grained execution control |
 
-**Note:** Step execution is currently supported by QwenImagePipeline only. See [Supported Models](#supported-models) for details.
+**Note:** Request-level batching is available for pipelines that declare the
+request-batch forward contract. Step execution is currently supported by
+QwenImagePipeline only. See [Supported Models](#supported-models) for details.
 
 ### Quantization Methods
 
@@ -122,14 +126,14 @@ The following tables show which models support each feature:
 | **MagiHuman**            |     ❌     |     ❌      |           ❌           |       ❓        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **MammothModa2(T2I)**    |     ❌     |     ❌      |           ❌           |       ❌        |         ❌         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Nextstep_1(T2I)**      |     ❓     |     ❓      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
-| **OmniGen2**             |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **OmniGen2**             |     ❌     |     ✅      |           ✅           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Ovis-Image**           |     ❌     |     ✅      |           ❌           |       ✅        |         ❌         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Qwen-Image**           |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ✅        |        ✅         |
 | **Qwen-Image-2512**      |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ✅        |        ✅         |
 | **Qwen-Image-Edit**      |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Qwen-Image-Edit-2509** |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |        ✅ (decode)         |          ✅           |       ❌        |        ❌         |
 | **Qwen-Image-Layered**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
-| **SenseNova-U1**         |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **SenseNova-U1**         |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Stable-Diffusion-XL**  |     ❌     |     ❌      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Stable-Diffusion3.5**  |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Z-Image**              |     ✅     |     ✅      |           ✅           |       ❓        |   ✅ (TP=2 only)   |          ❌          |   ✅    |             ❌             |      ✅ (decode)      |       ✅        |        ❌         |
@@ -146,7 +150,7 @@ The following tables show which models support each feature:
 | Model                        | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | Pipeline-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
 |------------------------------|:---------:|:----------:|:---------------------:|:--------------:|:-----------------:|:-----------------:|:------:|:-------------------------:|:--------------------:|:--------------:|:----------------:|
 | **Wan2.2**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ✅         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ❌        |        ❌         |
-| **Wan2.2-S2V**               |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ❌    |             ✅             |  ✅ (encode/decode)   |       ❌        |        ❌         |
+| **Wan2.2-S2V**               |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ❌        |        ❌         |
 | **Wan2.1-VACE**              |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ❌    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
@@ -270,7 +274,7 @@ Measured on NVIDIA H800:
 **Memory Optimization:**
 
 - **[CPU Offload Guide](diffusion/cpu_offload_diffusion.md)** - Offload model components to CPU, reduce GPU memory usage
-- **[VAE Patch Parallelism Guide](diffusion/parallelism/vae_patch_parallel.md)** - Distribute VAE decode tiling across GPUs for high-resolution images
+- **[VAE Parallelism Guide](diffusion/parallelism/vae_parallelism.md)** - Distribute VAE decode work across GPUs for high-resolution images and videos
 - **[Quantization Overview](quantization/overview.md)** - Overview of quantization methods for diffusion, multi-stage omni/TTS, and multi-stage diffusion models
 
 **Extensions:**

docs/user_guide/examples/offline_inference/image_to_video.md

@@ -3,7 +3,9 @@
 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_video>.
 
 
-This example demonstrates how to generate videos from images using Wan2.2 Image-to-Video models with vLLM-Omni's offline inference API.
+This example demonstrates how to generate videos from images using Wan2.2,
+LTX-2/LTX-2.3, and HunyuanVideo-1.5 Image-to-Video models with vLLM-Omni's
+offline inference API.
 
 ## Local CLI Usage
 
@@ -51,13 +53,41 @@ python image_to_video.py \
   --output i2v_output.mp4
 ```
 
+### LTX-2.3
+
+```bash
+python image_to_video.py \
+  --model dg845/LTX-2.3-Diffusers \
+  --model-class-name LTX23ImageToVideoPipeline \
+  --image cherry_blossom.jpg \
+  --prompt "Cherry blossoms swaying gently in the breeze with synchronized ambient sound" \
+  --negative-prompt "worst quality, inconsistent motion, blurry, jittery, distorted" \
+  --height 384 \
+  --width 512 \
+  --num-frames 25 \
+  --guidance-scale 4.0 \
+  --num-inference-steps 20 \
+  --frame-rate 24 \
+  --fps 24 \
+  --output ltx23_i2v_output.mp4
+```
+
+Use a Diffusers-format checkpoint such as `dg845/LTX-2.3-Diffusers`; the
+upstream `Lightricks/LTX-2.3` raw safetensors repo is not directly loadable by
+this pipeline. Pass `--model-class-name LTX23ImageToVideoPipeline` to select
+the LTX-2.3 image-to-video pipeline.
+
 Key arguments:
 
-- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V).
+- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V, or
+  LTX-2/LTX-2.3).
 - `--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).
+- `--height/--width`: Output resolution (auto-calculated from image if not set).
+  Wan dimensions should be multiples of 16; LTX dimensions should be multiples
+  of 32.
+- `--num-frames`: Number of frames (model-specific default; LTX-style models
+  work best with `8k + 1`).
 - `--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.
@@ -65,6 +95,7 @@ Key arguments:
 - `--sample-solver`: Wan2.2 sampling solver. Use `unipc` for the default multistep solver, or `euler` for Lightning/Distill checkpoints.
 - `--num-inference-steps`: Number of denoising steps (default 50).
 - `--fps`: Frames per second for the saved MP4 (requires `diffusers` export_to_video).
+- `--audio-sample-rate`: fallback audio sample rate for embedded audio.
 - `--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.

docs/user_guide/examples/offline_inference/mammothmodal2_preview.md

@@ -7,29 +7,22 @@ Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inf
 
 Download model
 ```bash
-hf bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-Preview
+hf download bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-Preview
 ```
 
 ### Text-to-Image (T2I)
 
-```bash
-python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_t2i.py \
-  --model ./MammothModa2-Preview \
-  --stage-config ./vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml \
-  --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
-  --height 1024 \
-  --width 1024 \
-  --num-inference-steps 50 \
-  --text-guidance-scale 4.0 \
-  --out output.png
-```
+Text-to-image now runs through the shared offline image example
+(`examples/offline_inference/text_to_image/text_to_image.py`). See the recipe
+`recipes/MammothModa2/MammothModa2-Preview.md` for the full command and the
+`extra_body` knobs (`text_guidance_scale`, `cfg_range`, `num_inference_steps`).
 
 ### Image Summary
 
 ```bash
 python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \
   --model ./MammothModa2-Preview \
-  --stage-config ./vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml \
+  --deploy-config ./vllm_omni/deploy/mammoth_moda2_ar.yaml \
   --question "Summarize this image." \
   --image ./image.png
 ```
@@ -40,7 +33,3 @@ python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_s
     ``````py
     --8<-- "examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py"
     ``````
-??? abstract "run_mammothmoda2_t2i.py"
-    ``````py
-    --8<-- "examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_t2i.py"
-    ``````

docs/user_guide/examples/offline_inference/text_to_image.md

@@ -160,9 +160,11 @@ python examples/offline_inference/text_to_image/text_to_image.py \
   --output flux2-dev.png
 ```
 
-### Batch Requests (Multiple Prompts)
+### Multiple Prompts
 
-You can pass multiple prompts in a single `generate` call.
+You can pass multiple prompts in a single `generate` call. For diffusion
+pipelines, each prompt is submitted as a separate logical request; compatible
+requests may be automatically batched by the scheduler and runner.
 
 ```python
 from vllm_omni.entrypoints.omni import Omni
@@ -181,17 +183,8 @@ if __name__ == "__main__":
 
 !!! info
 
-    Not all models support batch inference, and batch requesting mostly does not provide significant
-    performance improvement. This feature is primarily for interface compatibility with vLLM and to
-    allow for future improvements.
-
-!!! info
-
-    For diffusion pipelines, the stage config field `stage_args.[].runtime.max_batch_size` is 1 by
-    default, and the input list is sliced into single-item requests before feeding into the diffusion
-    pipeline. For models that do internally support batched inputs, you can
-    [modify this configuration](https://github.com/vllm-project/vllm-omni/tree/main/configuration/stage_configs.md) to let the model accept a
-    longer batch of prompts.
+    For diffusion request-level batching controls such as `max_num_seqs`, see
+    [Request-Level Batching](../../diffusion/request_batching.md).
 
 ### Negative Prompts
 

docs/user_guide/examples/offline_inference/text_to_video.md

@@ -12,6 +12,7 @@ For backend selection and SageAttention usage, see the [Diffusion Attention Back
 | Model | Default Resolution | Default Frames | Default Steps | Guidance | VRAM (BF16) |
 |---|---|---|---|---|---|
 | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 720x1280 | 81 | 40 | 4.0 | ~60 GiB |
+| `dg845/LTX-2.3-Diffusers` | 384x512 | 25 | 20 | 4.0 | 96GB-class GPU |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | 480x832 | 121 | 50 | 6.0 | 1×A100 80GB |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
 
@@ -50,6 +51,30 @@ python text_to_video.py \
   --output ltx2_out.mp4
 ```
 
+### LTX-2.3
+
+```bash
+python text_to_video.py \
+  --model dg845/LTX-2.3-Diffusers \
+  --model-class-name LTX23Pipeline \
+  --prompt "Cherry blossoms swaying gently in the breeze with synchronized ambient sound" \
+  --negative-prompt "worst quality, inconsistent motion, blurry, jittery, distorted" \
+  --height 384 \
+  --width 512 \
+  --num-frames 25 \
+  --num-inference-steps 20 \
+  --guidance-scale 4.0 \
+  --frame-rate 24 \
+  --fps 24 \
+  --audio-sample-rate 48000 \
+  --output ltx23_t2v_output.mp4
+```
+
+Use a Diffusers-format checkpoint such as `dg845/LTX-2.3-Diffusers`; the
+upstream `Lightricks/LTX-2.3` raw safetensors repo is not directly loadable by
+this pipeline. Pass `--model-class-name LTX23Pipeline` to select the LTX-2.3
+text-to-video pipeline explicitly.
+
 ### HunyuanVideo-1.5 (480p)
 
 ```bash
@@ -110,6 +135,8 @@ python text_to_video.py \
 ### Common
 
 - `--model`: Diffusers model ID or local path.
+- `--model-class-name`: Optional explicit pipeline class. Use `LTX23Pipeline`
+  for LTX-2.3 text-to-video.
 - `--prompt`: text description (string).
 - `--height/--width`: output resolution. Default depends on model.
 - `--num-frames`: number of frames. Default depends on model.
@@ -124,7 +151,8 @@ python text_to_video.py \
 - `--enable-cpu-offload`: enable CPU offloading for diffusion models.
 - `--enable-layerwise-offload`: enable layerwise offloading on DiT modules.
 - `--frame-rate`: generation FPS for pipelines that require it (e.g., LTX2).
-- `--audio-sample-rate`: audio sample rate for embedded audio (when the pipeline returns audio).
+- `--audio-sample-rate`: audio sample rate for embedded audio (when the
+  pipeline returns audio; LTX-2.3 outputs 48kHz audio).
 - `--quantization`: quantization method (`fp8` for FP8, `gguf` for GGUF).
 - `--flow-shift`: scheduler flow_shift parameter.
 

docs/user_guide/examples/online_serving/diffusers_pipeline_adapter.md

@@ -9,7 +9,8 @@ vLLM-Omni supports running diffusion models with the diffusers backend, directly
 
 The diffusers backend is a black-box adapter. Its primary focus is to serve diffusion models online.
 Currently, the following features are NOT yet supported.
-It is not guaranteed whether they will be supported in the future.
+Community contributions are welcome to add these features when they can be
+implemented with clear Diffusers or vLLM-Omni behavior.
 
 - CFG parallel execution
 - Sequence parallel execution
@@ -66,6 +67,35 @@ This is suitable for sampling parameters not available through the vLLM-Omni int
 When a parameter is available in the vLLM-Omni interface, it will be adapted here.
 But if that parameter is simultaneously set in both the vLLM-Omni interface and `diffusers_call_kwargs`, the **former** will take precedence (because it is set at request time).
 
+### Quantization
+
+Use `diffusers_load_kwargs` for Diffusers-native quantization options. If
+`diffusers_load_kwargs["quantization_config"]` is provided as a dictionary, the
+diffusers backend builds a Diffusers quantization config and lets Diffusers
+validate it before calling `DiffusionPipeline.from_pretrained()`.
+
+The diffusers backend also provides a small compatibility shortcut for
+vLLM-Omni quantization configs when `diffusers_load_kwargs` does not already
+contain `quantization_config`. Currently this maps online/dynamic `fp8` and
+online/dynamic `int8` to Diffusers/TorchAO dynamic quantization for transformer
+components such as `transformer` or `transformer_2`. This path requires
+`torchao` to be installed.
+
+For example, the CLI can request dynamic FP8 through the vLLM-Omni interface:
+
+```bash
+vllm serve "Qwen/Qwen-Image" \
+    --omni \
+    --diffusion-load-format diffusers \
+    --quantization-config '{"method": "fp8"}'
+```
+
+Other vLLM-Omni quantization methods, such as `gguf`, `modelopt`, `mxfp4`,
+`mxfp8`, serialized checkpoints, static FP8 configs, and layer-name skip lists
+such as `ignored_layers`, are intentionally not translated by this compatibility
+shortcut. Use Diffusers-native configuration through `diffusers_load_kwargs` or
+a native vLLM-Omni pipeline for those cases.
+
 ### Attention Backends
 
 The diffusers backend converts

docs/user_guide/examples/online_serving/text_to_speech.md

@@ -79,7 +79,7 @@ response = client.audio.speech.create(
 response.stream_to_file("output.wav")
 ```
 
-Streaming PCM output (where supported) — set `stream=true` with `response_format="pcm"`:
+Streaming PCM output (where supported) — set `stream=true`, `stream_format="audio"`, and `response_format="pcm"`:
 
 ```bash
 curl -X POST http://localhost:8091/v1/audio/speech \
@@ -88,6 +88,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "input": "Hello, how are you?",
         "voice": "default",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
@@ -414,10 +415,11 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "voice": "vivian",
         "language": "English",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
-Streaming requires `response_format="pcm"` and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
+Raw PCM streaming requires `stream_format="audio"`, `response_format="pcm"`, and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
 
 ### Streaming WebSocket
 The `/v1/audio/speech/stream` endpoint accepts text incrementally, splits it at sentence boundaries, and emits one PCM stream per sentence:
@@ -496,7 +498,7 @@ After startup, `/v1/audio/voices` lists `alice`, and `/v1/audio/speech` can use
 ```bash
 python voxcpm2/gradio_demo.py
 ```
-Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Audio is streamed from the OpenAI Speech endpoint with `stream=true`.
+Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Raw PCM audio is streamed from the OpenAI Speech endpoint with `stream=true` and `stream_format="audio"`.
 
 ---
 

docs/user_guide/quantization/fp8.md

@@ -44,10 +44,16 @@ pip install vllm-omni[quack]
 pip install 'quack-kernels[cu13]' --extra-index-url https://download.pytorch.org/whl/cu130
 ```
 
-It is enabled automatically once installed (no flag needed) and is **Blackwell-only**:
-on Hopper/Ada the CUTLASS FP8 kernel already fuses bias, so quack is not used there.
-Set `VLLM_OMNI_USE_QUACK_FP8=0` to force the FlashInfer path. If `quack-kernels` is
-not installed, FP8 still works — it just keeps the unfused FlashInfer path.
+It is enabled automatically once installed (no flag needed) and is **datacenter
+Blackwell only** (`sm_100` / `sm_101` / `sm_103`, compute capability `10.x`, e.g.
+B200): quack's CuteDSL GEMM uses the 5th-gen `tcgen05` tensor-core MMA, which exists
+only on those parts. On Hopper/Ada the CUTLASS FP8 kernel already fuses bias, and on
+workstation/consumer Blackwell (`sm_120` / `sm_121`, compute capability `12.x`, e.g.
+RTX PRO 6000 / RTX 50-series) `tcgen05` is absent — so quack is **not** auto-enabled
+there and FlashInfer's native FP8 path is used instead. Set
+`VLLM_OMNI_USE_QUACK_FP8=1` to force quack on, or `VLLM_OMNI_USE_QUACK_FP8=0` to force
+the FlashInfer path. If `quack-kernels` is not installed, FP8 still works — it just
+keeps the unfused FlashInfer path.
 
 #### Compile cache and warmup
 

docs/user_guide/quantization/mxfp4.md

@@ -39,7 +39,7 @@ vLLM-Omni provides two quantization methods with different scale structures:
 | NVIDIA Blackwell GPU (SM 100+) | ⭕ |
 | NVIDIA Ada/Hopper GPU (SM 89+) | ⭕ |
 | NVIDIA Ampere GPU (SM 80+) | ⭕ |
-| AMD ROCm | ⭕ |
+| AMD ROCm (gfx950 / MI355X) | ✅ |
 | Intel XPU | ⭕ |
 | Ascend NPU (Atlas 950 A5) | ✅ |
 

docs/user_guide/quantization/overview.md

@@ -20,7 +20,7 @@ type has a different quantization scope.
 | NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
 | NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
 | NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
-| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ❌ |
+| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ⭕ | ⭕ | ❌ |
 | Intel XPU | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
 | Ascend NPU | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ |
 

examples/offline_inference/audiox/README.md

@@ -1,40 +0,0 @@
-# AudioX offline inference
-
-Generate audio with the [AudioX](https://zeyuet.github.io/AudioX/) MMDiT diffusion
-pipeline (`AudioXPipeline`). Six tasks: `t2a`, `t2m`, `v2a`, `v2m`, `tv2a`, `tv2m`.
-
-## Prerequisites
-
-Download a vLLM-Omni weight bundle (component-sharded safetensors):
-
-```bash
-huggingface-cli download zhangj1an/AudioX --local-dir ./audiox_weights
-```
-
-The Hugging Face id `zhangj1an/AudioX` also works directly without prefetching.
-
-## Usage
-
-```bash
-# Text-to-audio only (default uses zhangj1an/AudioX from the Hub):
-python end2end.py --tasks t2a
-
-# All six tasks against a local bundle and a sample video for v2*/tv2*:
-python end2end.py \
-  --model ./audiox_weights \
-  --video https://zeyuet.github.io/AudioX/static/samples/V2M/1XeBotOFqHA.mp4
-
-# Subset of tasks, custom seed and steps:
-python end2end.py --tasks t2a tv2a --num-inference-steps 100 --seed 0
-```
-
-## Arguments
-
-- `--model`: HF id or local bundle path (default: `zhangj1an/AudioX`).
-- `--tasks`: any subset of `t2a t2m v2a v2m tv2a tv2m` (default: all).
-- `--video`: video file/URL — required for `v2*` and `tv2*`.
-- `--reference-audio`: optional audio prompt (audio-conditioned generation).
-- `--num-inference-steps`, `--guidance-scale`, `--seed`, `--seconds-total`,
-  `--sample-rate`, `--output-dir`: generation knobs.
-
-Outputs land in `<output-dir>/<task>.wav` as 16-bit stereo WAV.

examples/offline_inference/audiox/end2end.py

@@ -1,110 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""End-to-end AudioX offline example covering the 6 t2*/v2*/tv2* tasks.
-
-Provide a directory with the **vLLM-Omni AudioX safetensors bundle** (e.g. from
-``zhangj1an/AudioX`` on Hugging Face)::
-
-    huggingface-cli download zhangj1an/AudioX --local-dir ./audiox_weights
-    python end2end.py --model ./audiox_weights
-    python end2end.py --model ./audiox_weights --tasks t2a tv2a
-"""
-
-from __future__ import annotations
-
-import argparse
-import time
-from pathlib import Path
-
-import soundfile
-import torch
-import torchaudio.functional as TF
-
-from vllm_omni.entrypoints.omni import Omni
-from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-from vllm_omni.platforms import current_omni_platform
-
-ROOT = Path(__file__).resolve().parent
-
-SAMPLE_PROMPTS: dict[str, str] = {
-    "t2a": "Fireworks burst twice, followed by a period of silence before a clock begins ticking.",
-    "t2m": "Uplifting ukulele tune for a travel vlog",
-    "v2a": "",
-    "v2m": "",
-    "tv2a": "drum beating sound and human talking",
-    "tv2m": "uplifting music matching the scene",
-}
-
-ALL_TASKS = ("t2a", "t2m", "v2a", "v2m", "tv2a", "tv2m")
-VIDEO_TASKS = frozenset({"v2a", "v2m", "tv2a", "tv2m"})
-TEXT_TASKS = frozenset({"t2a", "t2m", "tv2a", "tv2m"})
-
-
-def parse_args() -> argparse.Namespace:
-    p = argparse.ArgumentParser(description="AudioX offline end-to-end (6 t2*/v2*/tv2* tasks).")
-    p.add_argument("--model", default="zhangj1an/AudioX", help="HF id or local AudioX bundle path.")
-    p.add_argument("--tasks", nargs="+", default=list(ALL_TASKS), choices=ALL_TASKS)
-    p.add_argument("--video", default="", help="Video path / URL (required for v2*/tv2*).")
-    p.add_argument("--reference-audio", default="", help="Optional audio prompt for audio-conditioned generation.")
-    p.add_argument("--output-dir", default=str(ROOT / "audiox_task_outputs"))
-    p.add_argument("--num-inference-steps", type=int, default=250)
-    p.add_argument("--seconds-total", type=float, default=10.0)
-    p.add_argument("--guidance-scale", type=float, default=6.0)
-    p.add_argument("--seed", type=int, default=42)
-    p.add_argument("--sample-rate", type=int, default=48000, help="Output WAV rate (resampled if != model rate).")
-    return p.parse_args()
-
-
-def save_wav(audio: torch.Tensor, path: Path, sample_rate: int) -> None:
-    """Write 16-bit PCM WAV. ``audio`` is ``[channels, samples]`` float in [-1, 1]."""
-    path.parent.mkdir(parents=True, exist_ok=True)
-    soundfile.write(str(path), audio.clamp(-1.0, 1.0).cpu().T.numpy(), sample_rate, subtype="PCM_16")
-
-
-def main() -> None:
-    args = parse_args()
-
-    omni = Omni(model=args.model, model_class_name="AudioXPipeline")
-
-    for task in args.tasks:
-        if task in VIDEO_TASKS and not args.video:
-            raise SystemExit(f"task={task!r} requires --video")
-        prompt = SAMPLE_PROMPTS[task] if task in TEXT_TASKS else ""
-        extra: dict = {"audiox_task": task, "seconds_start": 0.0, "seconds_total": float(args.seconds_total)}
-        if task in VIDEO_TASKS:
-            extra["video_path"] = args.video
-        if args.reference_audio:
-            extra["audio_path"] = args.reference_audio
-
-        generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(args.seed)
-        t0 = time.perf_counter()
-        outputs = omni.generate(
-            prompt,
-            OmniDiffusionSamplingParams(
-                generator=generator,
-                guidance_scale=args.guidance_scale,
-                num_inference_steps=args.num_inference_steps,
-                seed=args.seed,
-                extra_args=extra,
-            ),
-        )
-        audio = outputs[0].request_output.multimodal_output.get("audio")
-        if audio is None:
-            raise RuntimeError(f"No audio produced for task {task!r}")
-        audio = torch.as_tensor(audio).detach().cpu().float()
-        if audio.ndim == 3:
-            audio = audio[0]
-
-        model_sr = int(outputs[0].request_output.multimodal_output.get("audio_sample_rate") or 44100)
-        if model_sr != args.sample_rate:
-            audio = TF.resample(audio, model_sr, args.sample_rate)
-
-        out_path = Path(args.output_dir) / f"{task}.wav"
-        save_wav(audio, out_path, args.sample_rate)
-        print(f"[{task}] saved {out_path} ({time.perf_counter() - t0:.2f}s)")
-
-    omni.close()
-
-
-if __name__ == "__main__":
-    main()

examples/offline_inference/custom_pipeline/image_to_image/custom_pipeline.py

@@ -2,14 +2,12 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import logging
-from typing import Any
 
-import PIL.Image
 import torch
 
-from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
+from vllm_omni.diffusion.data import OmniDiffusionConfig
 from vllm_omni.diffusion.models.qwen_image.pipeline_qwen_image_edit import QwenImageEditPipeline
-from vllm_omni.diffusion.request import OmniDiffusionRequest
+from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
 
 logger = logging.getLogger(__name__)
 
@@ -18,58 +16,14 @@ class CustomPipeline(QwenImageEditPipeline):
     def __init__(self, *, od_config: OmniDiffusionConfig, prefix: str = ""):
         super().__init__(od_config=od_config, prefix=prefix)
 
-    def forward(
-        self,
-        req: OmniDiffusionRequest,
-        prompt: str | list[str] | None = None,
-        negative_prompt: str | list[str] | None = None,
-        image: PIL.Image.Image | torch.Tensor | None = None,
-        true_cfg_scale: float = 4.0,
-        height: int | None = None,
-        width: int | None = None,
-        num_inference_steps: int = 50,
-        sigmas: list[float] | None = None,
-        guidance_scale: float = 1.0,
-        num_images_per_prompt: int = 1,
-        generator: torch.Generator | list[torch.Generator] | None = None,
-        latents: torch.Tensor | None = None,
-        prompt_embeds: torch.Tensor | None = None,
-        prompt_embeds_mask: torch.Tensor | None = None,
-        negative_prompt_embeds: torch.Tensor | None = None,
-        negative_prompt_embeds_mask: torch.Tensor | None = None,
-        output_type: str | None = "pil",
-        attention_kwargs: dict[str, Any] | None = None,
-        callback_on_step_end_tensor_inputs: list[str] = ["latents"],
-        max_sequence_length: int = 512,
-    ) -> DiffusionOutput:
+    def forward(self, req: DiffusionRequestBatch):
         """Forward pass for image editing with dummy trajectory data."""
-        # Call parent's forward to get the normal output
-        output = super().forward(
-            req=req,
-            prompt=prompt,
-            negative_prompt=negative_prompt,
-            image=image,
-            true_cfg_scale=true_cfg_scale,
-            height=height,
-            width=width,
-            num_inference_steps=num_inference_steps,
-            sigmas=sigmas,
-            guidance_scale=guidance_scale,
-            num_images_per_prompt=num_images_per_prompt,
-            generator=generator,
-            latents=latents,
-            prompt_embeds=prompt_embeds,
-            prompt_embeds_mask=prompt_embeds_mask,
-            negative_prompt_embeds=negative_prompt_embeds,
-            negative_prompt_embeds_mask=negative_prompt_embeds_mask,
-            output_type=output_type,
-            attention_kwargs=attention_kwargs,
-            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
-            max_sequence_length=max_sequence_length,
-        )
+        # Customize num_inference_steps that best suits your model
+        actual_num_steps = req.sampling_params.num_inference_steps or 50
+        req.sampling_params.num_inference_steps = actual_num_steps
 
-        # Get actual num_inference_steps used
-        actual_num_steps = req.sampling_params.num_inference_steps or num_inference_steps
+        # Call parent's forward to get the normal output
+        output = super().forward(req=req)
 
         # Create dummy trajectory data
         dummy_trajectory_latents = torch.randn(actual_num_steps, 1, 16, 64, 64, dtype=torch.float32)

examples/offline_inference/image_to_image/image_edit.py

@@ -138,6 +138,15 @@ def parse_args() -> argparse.Namespace:
         required=True,
         help="Path(s) to input image file(s) (PNG, JPG, etc.). Can specify multiple images.",
     )
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help=(
+            "Path to a deploy YAML. Required for multi-stage image-edit pipelines "
+            "whose deploy config is not auto-loaded."
+        ),
+    )
     parser.add_argument(
         "--prompt",
         type=str,
@@ -349,7 +358,12 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--enforce-eager",
         action="store_true",
-        help="Disable torch.compile and force eager execution.",
+        default=None,
+        help=(
+            "Disable torch.compile and force eager execution. Left unset (None) "
+            "so it is only forwarded when explicitly given; "
+            "otherwise the per-stage deploy YAML value wins."
+        ),
     )
     parser.add_argument(
         "--vae-use-slicing",
@@ -468,7 +482,7 @@ def main():
         }
 
     # Initialize Omni with appropriate pipeline
-    omni = Omni(
+    omni_kwargs: dict[str, Any] = dict(
         model=args.model,
         enable_layerwise_offload=args.enable_layerwise_offload,
         vae_use_slicing=args.vae_use_slicing,
@@ -476,11 +490,15 @@ def main():
         cache_backend=args.cache_backend,
         cache_config=cache_config,
         parallel_config=parallel_config,
-        enforce_eager=args.enforce_eager,
         enable_cpu_offload=args.enable_cpu_offload,
         enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,
         profiler_config=args.profiler_config,
     )
+    if args.enforce_eager is not None:
+        omni_kwargs["enforce_eager"] = args.enforce_eager
+    if args.deploy_config:
+        omni_kwargs["deploy_config"] = args.deploy_config
+    omni = Omni(**omni_kwargs)
     model_class_name = get_model_class_name(omni)
     declared_extra_body_params = get_extra_body_params(model_class_name)
     print("Pipeline loaded")

examples/offline_inference/image_to_video/README.md

@@ -1,16 +1,142 @@
 # Image-To-Video
 
-This shared example generates videos from images with VACE, Wan2.2, LTX2,
-HunyuanVideo-1.5, Cosmos3, and other compatible pipelines.
+This shared example generates videos from images with VACE, Wan2.2,
+LTX-2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, and other compatible pipelines.
 
-## Local CLI Usage
+- `image_to_video.py`: command-line script for single video generation with advanced options.
 
-Download the example image:
+## Table of Contents
+
+- [Overview](#overview)
+- [Prerequisites](#prerequisites)
+- [Quick Start](#quick-start)
+- [Key Arguments](#key-arguments)
+- [More CLI Examples](#more-cli-examples)
+- [Advanced Features](#advanced-features)
+- [FAQ](#faq)
+
+## Overview
+
+This folder provides a unified CLI script for image-to-video generation using vLLM-Omni diffusion/video pipelines. The script selects practical defaults for supported model families while still exposing common sampling, memory, and parallelism options.
+
+### Supported Models
+
+| Model | Default Resolution | Default Frames | Default Steps | Guidance | VRAM Notes |
+| ----- | ------------------ | -------------- | ------------- | -------- | ---------- |
+| `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480 x 832 | 81 | 50 | 5.0 | Around 60 GiB BF16 for basic single-card usage |
+| `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 480 x 832 | 81 | 50 | 4.0 | Around 20–25 GiB BF16, smallest I2V model |
+| `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v` | 480 x 832 | 121 | 50 | 6.0 | Around 100 GiB at default settings; the example enables `--enable-cpu-offload` + VAE tiling/slicing to fit an 80 GiB card |
+| LTX2 (local path + `--model-class-name LTX2ImageToVideoPipeline`) | 512 x 768 | 121 | 40 | 4.0 | Memory use depends on frame count and tensor parallelism |
+
+!!! info
+    Peak VRAM: based on basic single-card usage, batch size = 1, without any acceleration/optimization features. Some model weights cannot fit into one card with 80 GiB VRAM, which may need to use CPU offloading.
+
+Default model: `Wan-AI/Wan2.2-I2V-A14B-Diffusers`.
+
+## Prerequisites
+
+Download the example image used in the snippets below:
 
 ```bash
-wget https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/cherry_blossom.jpg
+curl -L -o cherry_blossom.jpg https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/cherry_blossom.jpg
 ```
 
+## Quick Start
+
+### Python API
+
+Single-prompt generation using TI2V-5B (lightweight):
+
+```python
+import PIL.Image
+import torch
+
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+if __name__ == "__main__":
+    image = PIL.Image.open("cherry_blossom.jpg").convert("RGB")
+    image = image.resize((576, 320))
+
+    omni = Omni(
+        model="Wan-AI/Wan2.2-TI2V-5B-Diffusers",
+        flow_shift=12.0,
+    )
+    outputs = omni.generate(
+        {
+            "prompt": "Cherry blossoms swaying gently in the breeze, petals falling",
+            "multi_modal_data": {"image": image},
+        },
+        OmniDiffusionSamplingParams(
+            height=320,
+            width=576,
+            num_frames=17,
+            num_inference_steps=20,
+            guidance_scale=4.0,
+            generator=torch.Generator(device="cuda").manual_seed(42),
+        ),
+    )
+    from diffusers.utils import export_to_video
+    frames = outputs[0].request_output.images
+    export_to_video(frames, "quick_test_i2v.mp4", fps=16)
+```
+
+### Local CLI Usage
+
+Quick test using TI2V-5B with a small resolution and few frames:
+
+```bash
+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" \
+  --height 320 \
+  --width 576 \
+  --num-frames 17 \
+  --guidance-scale 4.0 \
+  --num-inference-steps 20 \
+  --flow-shift 12.0 \
+  --fps 16 \
+  --output quick_test_i2v.mp4
+```
+
+## Key Arguments
+
+| Argument | Type | Default | Description |
+| -------- | ---- | ------- | ----------- |
+| `--model` | str | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | Diffusers I2V model ID or local path |
+| `--model-class-name` | str | `None` | Override model class name (e.g., `LTX2ImageToVideoPipeline`) |
+| `--image` | str | (required) | Path to input image |
+| `--prompt` | str | `""` | Text description of desired motion/animation |
+| `--negative-prompt` | str | `""` | Optional list of artifacts to suppress |
+| `--seed` | int | `42` | Random seed for deterministic sampling |
+| `--guidance-scale` | float | `5.0` | CFG scale |
+| `--guidance-scale-high` | float | `None` | Separate CFG for high-noise stage (MoE only) |
+| `--height` | int | auto | Video height (auto-calculated from image if not set). Multiples of 16 |
+| `--width` | int | auto | Video width (auto-calculated from image if not set). Multiples of 16 |
+| `--num-frames` | int | `81` | Number of frames |
+| `--num-inference-steps` | int | `50` | Number of denoising steps |
+| `--boundary-ratio` | float | `0.875` | Boundary split ratio for two-stage MoE models |
+| `--flow-shift` | float | `5.0` | Scheduler flow shift (5.0 for 720p, 12.0 for 480p) |
+| `--sample-solver` | str | `unipc` | Wan2.2 sampling solver (`unipc` or `euler` for Lightning/Distill) |
+| `--fps` | int | `None` | Frames per second for the saved MP4 |
+| `--frame-rate` | float | `None` | Generation frame rate for pipelines that require it (e.g., LTX2) |
+| `--output` | str | `i2v_output.mp4` | Path to save the generated video |
+| `--vae-use-slicing` | flag | off | Enable VAE slicing for memory optimization |
+| `--vae-use-tiling` | flag | off | Enable VAE tiling for memory optimization |
+| `--enable-cpu-offload` | flag | off | Enable CPU offloading for diffusion models |
+| `--enable-layerwise-offload` | flag | off | Enable layerwise offloading on DiT modules |
+| `--cfg-parallel-size` | int | `1` | Set to `2` to enable CFG Parallel |
+| `--tensor-parallel-size` | int | `1` | Tensor parallel size (effective for models that support TP, e.g. LTX2) |
+| `--ulysses-degree` | int | `1` | Ulysses sequence parallel degree |
+| `--ring-degree` | int | `1` | Ring sequence parallel degree |
+| `--cache-backend` | str | `None` | Cache backend: `cache_dit` or `tea_cache` |
+| `--use-hsdp` | flag | off | Enable Hybrid Sharded Data Parallel |
+| `--hsdp-shard-size` | int | `-1` | GPUs per shard group (-1 auto-calculates) |
+| `--hsdp-replicate-size` | int | `1` | Number of replica groups for HSDP |
+
+## More CLI Examples
+
 ### Wan2.2-I2V-A14B-Diffusers (MoE)
 
 ```bash
@@ -18,7 +144,7 @@ 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 "low quality, blurry" \
   --height 480 \
   --width 832 \
   --num-frames 48 \
@@ -28,7 +154,7 @@ python image_to_video.py \
   --boundary-ratio 0.875 \
   --flow-shift 12.0 \
   --fps 16 \
-  --output i2v_output.mp4
+  --output i2v_wan_moe.mp4
 ```
 
 ### Wan2.2-TI2V-5B-Diffusers (Unified)
@@ -38,7 +164,7 @@ 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 "low quality, blurry" \
   --height 480 \
   --width 832 \
   --num-frames 48 \
@@ -46,9 +172,91 @@ python image_to_video.py \
   --num-inference-steps 40 \
   --flow-shift 12.0 \
   --fps 16 \
-  --output i2v_output.mp4
+  --output i2v_wan_ti2v.mp4
 ```
 
+### HunyuanVideo-1.5 I2V (480p)
+
+```bash
+python image_to_video.py \
+  --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v \
+  --image cherry_blossom.jpg \
+  --prompt "Cherry blossoms swaying gently in the breeze, petals falling, smooth motion" \
+  --height 480 \
+  --width 832 \
+  --num-frames 121 \
+  --guidance-scale 6.0 \
+  --flow-shift 5.0 \
+  --num-inference-steps 50 \
+  --fps 24 \
+  --enable-cpu-offload \
+  --vae-use-tiling \
+  --vae-use-slicing \
+  --output hunyuan_i2v.mp4
+```
+
+### LTX2 Image-to-Video
+
+```bash
+python image_to_video.py \
+  --model /path/to/LTX-2 \
+  --model-class-name LTX2ImageToVideoPipeline \
+  --image cherry_blossom.jpg \
+  --prompt "A cinematic dolly shot of cherry blossoms" \
+  --height 512 \
+  --width 768 \
+  --num-frames 121 \
+  --num-inference-steps 40 \
+  --guidance-scale 4.0 \
+  --frame-rate 24 \
+  --fps 24 \
+  --output ltx2_i2v.mp4
+```
+
+## Advanced Features
+
+### CFG Parallel
+
+Set `--cfg-parallel-size 2` to enable CFG Parallel for faster inference on multi-GPU setups.
+See more examples in the [cfg_parallel user guide](../../../docs/user_guide/diffusion/parallelism/cfg_parallel.md).
+
+### Cache Acceleration
+
+Use `--cache-backend cache_dit` for Cache-DiT acceleration or `--cache-backend tea_cache` for Timestep Embedding Aware Cache:
+
+```bash
+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" \
+  --cache-backend cache_dit \
+  --output i2v_cached.mp4
+```
+
+### LTX-2.3
+
+```bash
+python image_to_video.py \
+  --model dg845/LTX-2.3-Diffusers \
+  --model-class-name LTX23ImageToVideoPipeline \
+  --image cherry_blossom.jpg \
+  --prompt "Cherry blossoms swaying gently in the breeze with synchronized ambient sound" \
+  --negative-prompt "worst quality, inconsistent motion, blurry, jittery, distorted" \
+  --height 384 \
+  --width 512 \
+  --num-frames 25 \
+  --guidance-scale 4.0 \
+  --num-inference-steps 20 \
+  --frame-rate 24 \
+  --fps 24 \
+  --output ltx23_i2v_output.mp4
+```
+
+Use a Diffusers-format checkpoint such as `dg845/LTX-2.3-Diffusers`; the
+upstream `Lightricks/LTX-2.3` raw safetensors repo is not directly loadable by
+this pipeline. Pass `--model-class-name LTX23ImageToVideoPipeline` to select
+the LTX-2.3 image-to-video pipeline.
+
 ### Cosmos3
 
 ```bash
@@ -66,15 +274,19 @@ python image_to_video.py \
 
 Key arguments:
 
-- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V).
+- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V,
+  LTX-2/LTX-2.3, Cosmos3, or VACE).
 - `--image`: Path to the first-frame or source image.
 - `--last-image`: Optional last-frame condition for models such as VACE.
 - `--mask-image`: Optional inpainting mask. White pixels are regenerated and black pixels are preserved.
 - `--reference-image`: Optional reference image; repeat it to provide multiple references.
 - `--extra-body`: JSON object of model-specific generation params, filtered against the model's declared `extra_body_params` (see [`vllm_omni/model_extras`](../../../vllm_omni/model_extras)). Used by Cosmos3.
 - `--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: model-specific — Wan 81, LTX2 121, Cosmos3 189).
+- `--height/--width`: Output resolution (auto-calculated from image if not set).
+  Wan dimensions should be multiples of 16; LTX dimensions should be multiples
+  of 32.
+- `--num-frames`: Number of frames (model-specific default; LTX-style models
+  work best with `8k + 1`; Cosmos3 defaults to 189).
 - `--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.
@@ -82,6 +294,7 @@ Key arguments:
 - `--sample-solver`: Wan2.2 sampling solver. Use `unipc` for the default multistep solver, or `euler` for Lightning/Distill checkpoints.
 - `--num-inference-steps`: Number of denoising steps (default: model-specific — Wan 50, LTX2 40, Cosmos3 35).
 - `--fps`: Frames per second for the saved MP4 (requires `diffusers` export_to_video).
+- `--audio-sample-rate`: fallback audio sample rate for embedded audio.
 - `--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.
@@ -191,3 +404,11 @@ The VACE T2V command is documented in the shared
 
 For Wan2.2 LightX2V-converted local Diffusers directories and related LoRA
 assets, see the [LoRA guide](../../../docs/user_guide/diffusion/lora.md#wan22-lightx2v-offline-assembly).
+
+## FAQ
+
+**OOM errors**: Try using `--vae-use-slicing` and `--vae-use-tiling` to reduce memory usage. For very large models, add `--enable-cpu-offload` or `--enable-layerwise-offload`.
+
+**Auto-calculated resolution**: If `--height` and `--width` are not provided, the script calculates output dimensions from the input image while maintaining aspect ratio and targeting 480 x 832 area (or 512 x 768 for LTX2).
+
+**Wan2.2 MoE vs unified**: `I2V-A14B` is a larger Mixture-of-Experts model with separate low/high-noise DiT stages (use `--guidance-scale-high` and `--boundary-ratio`). `TI2V-5B` is a smaller unified T2V+I2V model that does not need these extra arguments.

examples/offline_inference/image_to_video/image_to_video.py

@@ -2,7 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 """
-Image-to-Video generation example using Wan2.2 I2V/TI2V models, LTX2, HunyuanVideo-1.5, or Wan2.1 VACE.
+Image-to-Video generation example using Wan2.2 I2V/TI2V models, LTX2/LTX-2.3,
+HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE.
 
 Supports:
 - Wan2.2-I2V-A14B-Diffusers: MoE model with CLIP image encoder
@@ -27,6 +28,13 @@ Usage:
         --num-frames 121 --num-inference-steps 40 --guidance-scale 4.0 \
         --frame-rate 24 --fps 24 --output ltx2_i2v.mp4
 
+    # LTX-2.3 image-to-video
+    python image_to_video.py --model dg845/LTX-2.3-Diffusers \
+        --model-class-name LTX23ImageToVideoPipeline \
+        --image input.jpg --prompt "A cinematic dolly shot of a boat" \
+        --height 384 --width 512 --num-frames 25 --num-inference-steps 20 \
+        --frame-rate 24 --fps 24 --output ltx23_i2v.mp4
+
     # HunyuanVideo-1.5 I2V (480p)
     python image_to_video.py --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v \
         --image input.jpg --prompt "A cat playing with yarn" \
@@ -79,17 +87,20 @@ parse_profiler_config = functools.partial(parse_json_object, flag_name="--profil
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
-        description="Generate a video from one or more images (Wan2.2, LTX2, HunyuanVideo-1.5, or Wan2.1 VACE)."
+        description=(
+            "Generate a video from one or more images "
+            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE)."
+        )
     )
     parser.add_argument(
         "--model",
         default="Wan-AI/Wan2.2-I2V-A14B-Diffusers",
-        help="Diffusers I2V model ID or local path (Wan2.2, HunyuanVideo-1.5, or Wan2.1 VACE).",
+        help="Diffusers I2V model ID or local path (Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE).",
     )
     parser.add_argument(
         "--model-class-name",
         default=None,
-        help="Override model class name (e.g., LTX2ImageToVideoPipeline).",
+        help="Override model class name (e.g., LTX2ImageToVideoPipeline or LTX23ImageToVideoPipeline).",
     )
     parser.add_argument("--image", help="Path to the first-frame or source image.")
     parser.add_argument("--last-image", help="Path to a last-frame condition (used by models such as VACE).")
@@ -113,7 +124,10 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--width", type=int, default=None, help="Video width (auto-calculated from image if not set).")
     parser.add_argument("--num-frames", type=int, default=None, help="Number of frames. Default: model-specific.")
     parser.add_argument(
-        "--num-inference-steps", type=int, default=None, help="Sampling steps. Default: model-specific."
+        "--num-inference-steps",
+        type=int,
+        default=None,
+        help="Sampling steps. Default: model-specific.",
     )
     parser.add_argument("--boundary-ratio", type=float, default=0.875, help="Boundary split ratio for MoE models.")
     parser.add_argument(
@@ -312,10 +326,8 @@ def main():
     generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(args.seed)
     model_name = str(args.model).lower() if args.model is not None else ""
     model_class_name = args.model_class_name
-    is_ltx2 = "ltx2" in model_name or (model_class_name and "ltx2" in model_class_name.lower())
+    is_ltx2 = model_class_name in {"LTX2ImageToVideoPipeline", "LTX23ImageToVideoPipeline"}
     is_cosmos = "cosmos" in model_name or (model_class_name is not None and "cosmos" in model_class_name.lower())
-    if model_class_name is None and is_ltx2:
-        model_class_name = "LTX2ImageToVideoPipeline"
 
     image = PIL.Image.open(args.image).convert("RGB") if args.image else None
     last_image = PIL.Image.open(args.last_image).convert("RGB") if args.last_image else None
@@ -340,7 +352,15 @@ def main():
             16,
         )
     elif is_ltx2:
-        d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = 24, 4.0, 121, 40, 5.0, 512 * 768, 32
+        d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = (
+            24,
+            4.0,
+            121,
+            40,
+            5.0,
+            512 * 768,
+            32,
+        )
     else:  # Wan2.2 / HunyuanVideo-1.5
         d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = 16, 5.0, 81, 50, 5.0, 480 * 832, 16
 
@@ -462,7 +482,7 @@ def main():
         f" tensor_parallel_size={args.tensor_parallel_size}, vae_patch_parallel_size={args.vae_patch_parallel_size},"
         f" pipeline_parallel_size={args.pipeline_parallel_size}"
     )
-    print(f"  Video size: {args.width}x{args.height}")
+    print(f"  Video size: {width}x{height}")
     print(f"{'=' * 60}\n")
 
     prompt_dict = build_image_to_video_prompt(
@@ -509,6 +529,7 @@ def main():
     print(f"Total generation time: {generation_time:.4f} seconds ({generation_time * 1000:.2f} ms)")
 
     audio = None
+    audio_sample_rate = args.audio_sample_rate
     if isinstance(frames, list):
         frames = frames[0] if frames else None
 
@@ -519,11 +540,13 @@ def main():
             )
         if frames.multimodal_output and "audio" in frames.multimodal_output:
             audio = frames.multimodal_output["audio"]
+            audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
         if frames.is_pipeline_output and frames.request_output is not None:
             inner_output = frames.request_output
             if isinstance(inner_output, OmniRequestOutput):
                 if inner_output.multimodal_output and "audio" in inner_output.multimodal_output:
                     audio = inner_output.multimodal_output["audio"]
+                    audio_sample_rate = inner_output.multimodal_output.get("audio_sample_rate", audio_sample_rate)
                 frames = inner_output
         if isinstance(frames, OmniRequestOutput):
             if frames.images:
@@ -531,6 +554,7 @@ def main():
                     frames, audio = frames.images[0]
                 elif len(frames.images) == 1 and isinstance(frames.images[0], dict):
                     audio = frames.images[0].get("audio")
+                    audio_sample_rate = frames.images[0].get("audio_sample_rate", audio_sample_rate)
                     frames = frames.images[0].get("frames") or frames.images[0].get("video")
                 else:
                     frames = frames.images
@@ -543,6 +567,7 @@ def main():
             frames, audio = first_item
         elif isinstance(first_item, dict):
             audio = first_item.get("audio")
+            audio_sample_rate = first_item.get("audio_sample_rate", audio_sample_rate)
             frames = first_item.get("frames") or first_item.get("video")
         elif isinstance(first_item, list):
             frames = first_item
@@ -551,6 +576,7 @@ def main():
         frames, audio = frames
     elif isinstance(frames, dict):
         audio = frames.get("audio")
+        audio_sample_rate = frames.get("audio_sample_rate", audio_sample_rate)
         frames = frames.get("frames") or frames.get("video")
 
     if frames is None:
@@ -669,7 +695,7 @@ def main():
             frames_u8,
             audio_np,
             fps=float(fps),
-            audio_sample_rate=args.audio_sample_rate,
+            audio_sample_rate=audio_sample_rate,
         )
         with open(str(output_path), "wb") as f:
             f.write(video_bytes)

examples/offline_inference/mammothmodal2_preview/README.md

@@ -1,24 +1,17 @@
 # MammothModa2-Preview
 
-## Run examples (MammothModa2-Preview)
+> Text-to-image (T2I) generation has moved to the shared offline image example.
+> See the recipe: [`recipes/MammothModa2/MammothModa2-Preview.md`](../../../recipes/MammothModa2/MammothModa2-Preview.md).
+>
+> This directory retains the image-understanding (image → text) example below,
+> which is outside the scope of the diffusion `model_extras` example migration
+> (#4548 / RFC #4539: AR / Omni understanding examples keep dedicated scripts).
+
+## Run example (MammothModa2-Preview)
 
 Download model
 ```bash
-hf bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-Preview
-```
-
-### Text-to-Image (T2I)
-
-```bash
-python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_t2i.py \
-  --model ./MammothModa2-Preview \
-  --stage-config ./vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml \
-  --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
-  --height 1024 \
-  --width 1024 \
-  --num-inference-steps 50 \
-  --text-guidance-scale 4.0 \
-  --out output.png
+hf download bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-Preview
 ```
 
 ### Image Summary
@@ -26,7 +19,7 @@ python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_t2i.py
 ```bash
 python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \
   --model ./MammothModa2-Preview \
-  --stage-config ./vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml \
+  --deploy-config ./vllm_omni/deploy/mammoth_moda2_ar.yaml \
   --question "Summarize this image." \
   --image ./image.png
 ```

examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py

@@ -4,7 +4,7 @@ Offline inference example: MammothModa2 image summarization (single AR stage).
 Example:
   uv run python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \
     --model path/to/MammothModa2-Preview \
-    --stage-config vllm_omni/model_executor/stage_configs/mammoth_moda2_ar.yaml \
+    --deploy-config vllm_omni/deploy/mammoth_moda2_ar.yaml \
     --image /path/to/input.jpg \
     --question "Please summarize the content of this image."
 """
@@ -27,9 +27,7 @@ DEFAULT_QUESTION = "Please summarize the content of this image."
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(description="MammothModa2 image summarization (offline, AR only).")
     parser.add_argument("--model", type=str, required=True, help="Path to model directory or model id.")
-    parser.add_argument(
-        "--stage-config", type=str, required=True, help="Path to stage config yaml (single-stage AR->text)."
-    )
+    parser.add_argument("--deploy-config", type=str, required=True, help="Path to deploy config yaml.")
     parser.add_argument("--image", type=str, required=True, help="Path to input image.")
     parser.add_argument("--question", type=str, default=DEFAULT_QUESTION, help="Question/instruction for the model.")
     parser.add_argument("--system", type=str, default=DEFAULT_SYSTEM, help="System prompt.")
@@ -73,7 +71,7 @@ def main() -> None:
 
     omni = Omni(
         model=args.model,
-        stage_configs_path=args.stage_config,
+        deploy_config=args.deploy_config,
         trust_remote_code=args.trust_remote_code,
         enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,
     )

examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_t2i.py

@@ -1,248 +0,0 @@
-"""
-Offline inference example for MammothModa2 Text-to-Image (T2I) generation.
-This script uses the vllm_omni.Omni pipeline with a multi-stage configuration.
-
-Workflow:
-1. Stage 0 (AR): Generates visual tokens and their corresponding hidden states.
-2. Stage 1 (DiT): Consumes the hidden states as conditions to perform diffusion
-   and VAE decoding to produce the final image.
-
-Example Usage:
-    uv run python examples/offline_inference/run_mammothmoda2_t2i.py \
-        --model path/to/MammothModa2-Preview \
-        --stage-config vllm_omni/model_executor/stage_configs/mammoth_moda2.yaml \
-        --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
-        --out output.png
-"""
-
-from __future__ import annotations
-
-import argparse
-import json
-import logging
-import os
-from pathlib import Path
-from typing import NamedTuple
-
-import torch
-from PIL import Image
-from vllm.sampling_params import SamplingParams
-
-from vllm_omni import Omni
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
-logger = logging.getLogger(__name__)
-
-# ---------------------------------------------------------------------------
-# Constants
-# ---------------------------------------------------------------------------
-_PATCH_SIZE = 16  # AR image grid patch size (pixels per token)
-
-
-class T2IGenConfig(NamedTuple):
-    eol_token_id: int
-    visual_token_start_id: int
-    visual_token_end_id: int
-    top_k: int  # AR sampling top-k (covers the full visual generation vocabulary)
-    # Qwen2.5-VL special vision tokens: <|image_pad|>, <|video_pad|>, <|vision_start|>, <|vision_end|>
-    visual_ids: list[int]
-
-
-def load_t2i_generation_config(model_dir: str) -> T2IGenConfig:
-    """Load T2I token IDs from t2i_generation_config.json and config.json.
-
-    Supports both local directory paths and HuggingFace Hub model IDs.
-    """
-    model_path = Path(model_dir)
-
-    def _read_json(filename: str) -> dict:
-        local = model_path / filename
-        if local.exists():
-            with local.open(encoding="utf-8") as f:
-                return json.load(f)
-        # Fall back to HuggingFace Hub when model_dir is a repo ID.
-        try:
-            from huggingface_hub import hf_hub_download
-        except ImportError as exc:
-            raise ImportError(
-                "huggingface_hub is required to load configs from HF Hub. Install it with: pip install huggingface_hub"
-            ) from exc
-        cached = hf_hub_download(repo_id=model_dir, filename=filename)
-        with open(cached, encoding="utf-8") as f:
-            return json.load(f)
-
-    gen_cfg = _read_json("t2i_generation_config.json")
-    llm_cfg = _read_json("config.json").get("llm_config", {})
-
-    return T2IGenConfig(
-        eol_token_id=int(gen_cfg["eol_token_id"]),
-        visual_token_start_id=int(gen_cfg["visual_token_start_id"]),
-        visual_token_end_id=int(gen_cfg["visual_token_end_id"]),
-        top_k=int(gen_cfg["top_k"]),
-        visual_ids=[
-            int(llm_cfg["image_token_id"]),
-            int(llm_cfg["video_token_id"]),
-            int(llm_cfg["vision_start_token_id"]),
-            int(llm_cfg["vision_end_token_id"]),
-        ],
-    )
-
-
-def parse_args() -> argparse.Namespace:
-    p = argparse.ArgumentParser(description="Run MammothModa2 T2I (AR -> DiT) with vLLM-Omni.")
-    p.add_argument("--model", type=str, required=True, help="Path to the model directory.")
-    p.add_argument("--stage-config", type=str, required=True, help="Path to the multi-stage YAML configuration.")
-    p.add_argument(
-        "--prompt",
-        type=str,
-        action="append",
-        default=None,
-        help=(
-            "Text prompt for image generation. Can be provided multiple times "
-            "to generate multiple images with shared height/width/CFG settings."
-        ),
-    )
-    p.add_argument("--height", type=int, default=1024, help="Output image height (must be a multiple of 16).")
-    p.add_argument("--width", type=int, default=1024, help="Output image width (must be a multiple of 16).")
-    p.add_argument("--num-inference-steps", type=int, default=50, help="Number of diffusion steps for the DiT stage.")
-    p.add_argument(
-        "--text-guidance-scale", type=float, default=9.0, help="Classifier-Free Guidance (CFG) scale for DiT."
-    )
-    p.add_argument(
-        "--cfg-range",
-        type=float,
-        nargs=2,
-        default=(0.0, 1.0),
-        help="Relative step range [start, end] where CFG is active.",
-    )
-    p.add_argument("--out", type=str, default="output.png", help="Path to save the generated image.")
-    p.add_argument("--trust-remote-code", action="store_true", help="Trust remote code when loading the model.")
-    args = p.parse_args()
-    if not args.prompt:
-        args.prompt = ["A stylish woman with sunglasses riding a motorcycle in NYC."]
-    return args
-
-
-def tensor_to_pil(image: torch.Tensor) -> Image.Image:
-    """Convert a normalized torch tensor [-1, 1] to a PIL Image."""
-    if image.ndim == 4:
-        image = image[0]
-    image = image.detach().to("cpu")
-    image = (image / 2 + 0.5).clamp(0, 1)
-    image = (image * 255).to(torch.uint8)
-    image = image.permute(1, 2, 0).contiguous().numpy()
-    return Image.fromarray(image)
-
-
-def _format_prompt(user_prompt: str, ar_width: int, ar_height: int) -> str:
-    """Build the AR-stage prompt string including the image grid header."""
-    return (
-        "<|im_start|>system\nYou are a helpful image generator.<|im_end|>\n"
-        f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
-        "<|im_start|>assistant\n"
-        f"<|image start|>{ar_width}*{ar_height}<|image token|>"
-    )
-
-
-def _collect_images(outputs: list) -> list[torch.Tensor]:
-    """Extract all image tensors produced by the final (DiT) stage."""
-    images: list[torch.Tensor] = []
-    for out in outputs:
-        ro_item = getattr(out, "request_output", out)
-        for completion in getattr(ro_item, "outputs", None) or []:
-            mm = getattr(completion, "multimodal_output", None)
-            if not isinstance(mm, dict) or "image" not in mm:
-                raise RuntimeError(f"Missing image in multimodal output: {mm}")
-            payload = mm["image"]
-            for tensor in payload if isinstance(payload, list) else [payload]:
-                if not isinstance(tensor, torch.Tensor):
-                    raise TypeError(f"Expected image tensor, got {type(tensor)}")
-                images.append(tensor)
-    return images
-
-
-def _save_images(images: list[torch.Tensor], out_path: str) -> list[str]:
-    """Save image tensors to disk.
-
-    Single image: written to *out_path* exactly.
-    Multiple images: suffixed as ``<base>_0<ext>``, ``<base>_1<ext>``, …
-    """
-    if not images:
-        raise RuntimeError("No images to save.")
-    base, ext = os.path.splitext(out_path)
-    ext = ext or ".png"
-    paths = []
-    for i, tensor in enumerate(images):
-        path = out_path if len(images) == 1 else f"{base}_{i}{ext}"
-        tensor_to_pil(tensor).save(path)
-        paths.append(path)
-    return paths
-
-
-def main() -> None:
-    args = parse_args()
-    os.makedirs(os.path.dirname(args.out) or ".", exist_ok=True)
-
-    if args.height <= 0 or args.width <= 0:
-        raise ValueError(f"Height and width must be positive, got {args.height}x{args.width}")
-    if args.height % _PATCH_SIZE != 0 or args.width % _PATCH_SIZE != 0:
-        raise ValueError(f"Height and width must be multiples of {_PATCH_SIZE}, got {args.height}x{args.width}")
-
-    ar_height = args.height // _PATCH_SIZE
-    ar_width = args.width // _PATCH_SIZE
-    gen_cfg = load_t2i_generation_config(args.model)
-    expected_grid_tokens = ar_height * (ar_width + 1)
-
-    logger.info("Initializing Omni pipeline...")
-    omni = Omni(model=args.model, stage_configs_path=args.stage_config, trust_remote_code=args.trust_remote_code)
-    try:
-        ar_sampling = SamplingParams(
-            temperature=1.0,
-            top_p=1.0,
-            top_k=gen_cfg.top_k,
-            max_tokens=max(1, expected_grid_tokens + 1),  # +1 for hidden state of eoi
-            detokenize=False,
-        )
-        dit_sampling = SamplingParams(
-            temperature=0.0,
-            top_p=1.0,
-            top_k=-1,
-            max_tokens=1,
-            detokenize=False,
-        )
-
-        additional_information = {
-            "omni_task": ["t2i"],
-            "ar_width": [ar_width],
-            "ar_height": [ar_height],
-            "eol_token_id": [gen_cfg.eol_token_id],
-            "visual_token_start_id": [gen_cfg.visual_token_start_id],
-            "visual_token_end_id": [gen_cfg.visual_token_end_id],
-            "image_height": [args.height],
-            "image_width": [args.width],
-            "num_inference_steps": [args.num_inference_steps],
-            "text_guidance_scale": [args.text_guidance_scale],
-            "cfg_range": [args.cfg_range[0], args.cfg_range[1]],
-            "visual_ids": gen_cfg.visual_ids,
-        }
-        inputs = [
-            {
-                "prompt": _format_prompt(p, ar_width, ar_height),
-                "additional_information": dict(additional_information),
-            }
-            for p in args.prompt
-        ]
-
-        logger.info("Starting generation...")
-        # omni.generate() returns a Generator; consume it to run the full pipeline.
-        outputs = list(omni.generate(inputs, [ar_sampling, dit_sampling]))
-
-        logger.info("Post-processing and saving image(s)...")
-        for path in _save_images(_collect_images(outputs), args.out):
-            logger.info(f"Saved: {path}")
-    finally:
-        omni.close()
-
-
-if __name__ == "__main__":
-    main()

examples/offline_inference/ming_flash_omni/README.md

@@ -90,14 +90,40 @@ Pass `--deploy-config /path/to/your_deploy.yaml` to any of the commands above to
 
 ### Image generation (text-to-image / img2img)
 
-The diffusion-stage image-generation path is currently only wired through
-the online OpenAI-compatible chat endpoint (`/v1/chat/completions` with
-`"modalities": ["image"]`). For payloads, optional knobs
-(`steps`/`cfg`/`seed`/`byte5_text`/`negative_prompt`), and the img2img
-reference-image flow, see the
-[image-gen section in the online-serving README](../../online_serving/ming_flash_omni/README.md#image-generation-text-to-image).
-`end2end.py` does not yet exercise the imagegen stage; that is tracked as
-follow-up work.
+Image generation is served through the standard task examples.
+The diffusion-stage knobs are declared centrally in `vllm_omni/model_extras/ming_flash_omni.py` and routed via `--extra-body`. With deploy yaml assigned properly, the model can run through the shared example scripts like any other diffusion model:
+
+Text-to-image (offline):
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+    --model Jonathan1909/Ming-flash-omni-2.0 \
+    --deploy-config vllm_omni/deploy/ming_flash_omni_image.yaml \
+    --prompt "Please draw a cute cat." \
+    --height 1024 \
+    --width 1024 \
+    --extra-body '{"steps": 30, "cfg": 2.0, "seed": 42}' \
+    --output ming_flash_omni_t2i.png
+```
+
+Image-to-image (offline):
+```bash
+# Reference image: figures/cases/person_gen_05.png from the upstream Ming repo
+# https://github.com/inclusionAI/Ming/blob/3954fcb880ff5e61ff128bcf7f1ec344d46a6fe3/examples/vllm_demo.py
+wget https://raw.githubusercontent.com/inclusionAI/Ming/3954fcb880ff5e61ff128bcf7f1ec344d46a6fe3/figures/cases/person_gen_05.png
+
+python examples/offline_inference/image_to_image/image_edit.py \
+    --model Jonathan1909/Ming-flash-omni-2.0 \
+    --deploy-config vllm_omni/deploy/ming_flash_omni_image.yaml \
+    --image person_gen_05.png \
+    --prompt "Put a pair of sunglasses on the person." \
+    --extra-args '{"steps": 30, "cfg": 2.0, "seed": 42}' \
+    --output ming_flash_omni_i2i.png
+```
+
+For the online path and the full knob list
+(`steps`/`cfg`/`seed`/`byte5_text`/`negative_prompt`/`height`/`width`), see the
+[image-generation section in the recipe](../../../recipes/inclusionAI/Ming-flash-omni-2.0.md#image-generation-text-to-image--img2img).
 
 ## Online serving
 

examples/offline_inference/qwen2_5_omni/README.md

@@ -60,3 +60,24 @@ If media file paths are not provided, the script will use default assets. Suppor
 - `mixed_modalities`: Audio + image + video
 - `use_audio_in_video`: Extract audio from video
 - `text`: Text-only query
+
+### Composable parallelism (strategy configs)
+
+You can shard or replicate a stage with a small composable-parallel
+`strategy.yaml` overlaid onto the **bundled default deploy config** -- no bespoke
+deploy YAML required. Supply the strategy with `--strategy-config` and any matching
+device layout via `--stage-overrides` (explicit kwargs forwarded to `Omni`).
+
+Both examples below require >= 3 GPUs (the thinker uses 2, talker + code2wav share a third):
+
+```bash
+# Tensor-parallel the thinker (TP=2)
+python end2end.py --query-type text --num-prompts 6 \
+    --strategy-config strategy_tp2.yaml \
+    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
+
+# Replicate the thinker across 2 engines (round-robin load balancing)
+python end2end.py --query-type text --num-prompts 6 \
+    --strategy-config strategy_stage_replica.yaml \
+    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
+```

examples/offline_inference/qwen2_5_omni/end2end.py

@@ -553,6 +553,39 @@ def parse_args():
         default=False,
         help="Use py_generator mode. The returned type of Omni.generate() is a Python Generator object.",
     )
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help="Optional explicit deploy YAML (otherwise the bundled default is used).",
+    )
+    parser.add_argument(
+        "--strategy-config",
+        type=str,
+        default=None,
+        help=(
+            "Optional composable-parallel strategy.yaml mapping role -> parallel axes "
+            "(e.g. tp / stage_replica). Overlaid onto the merged stage configs."
+        ),
+    )
+    parser.add_argument(
+        "--stage-overrides",
+        type=str,
+        default=None,
+        help=(
+            "Optional JSON of per-stage overrides applied on top of the default "
+            'deploy config, e.g. \'{"0": {"devices": "0,1"}}\' to give the '
+            "thinker a 2-GPU pool. Lets you run a strategy on the bundled default "
+            "deploy config without writing a bespoke deploy YAML."
+        ),
+    )
+    parser.add_argument(
+        "--omni-lb-policy",
+        type=str,
+        default=None,
+        choices=["random", "round-robin", "least-queue-length"],
+        help="StagePool load-balancer policy for replicated stages (orchestrator-level knob).",
+    )
     return parser.parse_args()
 
 

examples/offline_inference/qwen2_5_omni/strategy_stage_replica.yaml

@@ -0,0 +1,22 @@
+# Composable-parallel strategy for Qwen2.5-Omni.
+#
+# Replicate the thinker stage across 2 independent engines and let the omni
+# StagePool balance requests round-robin across them. The thinker's `devices`
+# entry acts as a per-replica template, so omni expands it to one GPU per replica.
+# The `routing: round_robin` below auto-derives the StagePool load-balancer policy,
+# so --omni-lb-policy is not needed.
+#
+# Rather than hand-write a bespoke deploy YAML, run on the bundled default deploy
+# config and pass the matching 2-GPU device template via --stage-overrides (explicit
+# kwargs): the thinker gets a two-GPU pool while talker + code2wav share a third
+# card. Requires >= 3 GPUs.
+#
+# Usage:
+#   python3 end2end.py --query-type text --num-prompts 6 \
+#       --strategy-config strategy_stage_replica.yaml \
+#       --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
+strategies:
+  thinker:
+    - axis: stage_replica
+      size: 2
+      routing: round_robin

examples/offline_inference/qwen2_5_omni/strategy_tp2.yaml

@@ -0,0 +1,19 @@
+# Composable-parallel strategy for Qwen2.5-Omni: tensor-parallel the thinker.
+#
+# Shard the thinker stage across 2 GPUs with tensor parallelism (TP=2). TP routing
+# is fixed (broadcast), so no `routing` / load-balancer policy is involved -- this
+# is the simplest axis and a good starting point for the mental model.
+#
+# The strategy sets tensor_parallel_size=2 on the thinker. Rather than hand-write a
+# bespoke deploy YAML, run on the bundled default deploy config and pass a matching
+# 2-GPU device layout via --stage-overrides (explicit kwargs), so the thinker gets a
+# two-GPU pool while talker + code2wav share a third card. Requires >= 3 GPUs.
+#
+# Usage:
+#   python3 end2end.py --query-type text --num-prompts 6 \
+#       --strategy-config strategy_tp2.yaml \
+#       --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
+strategies:
+  thinker:
+    - axis: tp
+      size: 2

examples/offline_inference/sensenova_u1/README.md

@@ -38,9 +38,9 @@ python end2end.py --modality text2text \
 python end2end.py --prompt "A futuristic cityscape at sunset" \
                   --width 2048 --height 1024 --think
 
-# Text-to-image with Cache-DiT acceleration
+# Text-to-image with cache acceleration (cache_dit or tea_cache)
 python end2end.py --prompt "A cute cat" \
-                  --cache-backend cache_dit \
+                  --cache-backend tea_cache
 ```
 
 > **Note**: Default configuration works on a single **NVIDIA A100 (80GB)** or **H100** GPU.
@@ -139,7 +139,7 @@ python end2end.py \
 | `--tensor-parallel-size` | 1 | Number of GPUs for tensor parallelism |
 | `--enforce-eager` | False | Disable torch.compile |
 | `--enable-cpu-offload` | False | Enable module-wise (sequential) CPU offload to reduce peak VRAM |
-| `--cache-backend` | None | Set to `cache_dit` for Cache-DiT acceleration |
+| `--cache-backend` | None | Set to `cache_dit` or `tea_cache` for image generation acceleration |
 | `--enable-cache-dit-summary` | False | Print Cache-DiT cache statistics after generation |
 
 ## Reducing GPU Memory Usage
@@ -211,4 +211,4 @@ The img2img counterpart lives at
 ## Online Serving
 
 For OpenAI-compatible API serving, see
-[`examples/online_serving/sensenova_u1/`](../../examples/online_serving/sensenova_u1/).
+[`examples/online_serving/sensenova_u1/`](../../online_serving/sensenova_u1/README.md).

examples/offline_inference/sensenova_u1/end2end.py

@@ -139,6 +139,12 @@ def parse_args():
         default=1,
         help="Number of GPUs for tensor parallelism.",
     )
+    parser.add_argument(
+        "--cfg-parallel-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for CFG parallelism.",
+    )
     parser.add_argument(
         "--enforce-eager",
         action="store_true",
@@ -158,7 +164,7 @@ def parse_args():
         "--cache-backend",
         type=str,
         default=None,
-        choices=["cache_dit"],
+        choices=["cache_dit", "tea_cache"],
         help="Cache backend to use for image generation acceleration.",
     )
     parser.add_argument(
@@ -194,6 +200,7 @@ def main():
         enable_cpu_offload=args.enable_cpu_offload,
         cache_backend=args.cache_backend,
         enable_cache_dit_summary=args.enable_cache_dit_summary,
+        cfg_parallel_size=args.cfg_parallel_size,
     )
 
     extra_args = {
@@ -237,6 +244,7 @@ def main():
     print(f"  Seed           : {args.seed}")
     print(f"  Think mode     : {args.think}")
     print(f"  TP size        : {args.tensor_parallel_size}")
+    print(f"  CFG size       : {args.cfg_parallel_size}")
     print(f"  Cache backend  : {args.cache_backend or 'none'}")
     print(f"{'=' * 60}\n")
 

examples/offline_inference/speech_to_video/speech_to_video.py

@@ -172,6 +172,29 @@ def parse_args() -> argparse.Namespace:
         default=1,
         help="Number of GPUs used for ulysses sequence parallelism.",
     )
+    parser.add_argument(
+        "--use-hsdp",
+        action="store_true",
+        help="Enable Hybrid Sharded Data Parallel to shard model weights across GPUs.",
+    )
+    parser.add_argument(
+        "--hsdp-shard-size",
+        type=int,
+        default=-1,
+        help=(
+            "Number of GPUs to shard model weights across within each replica group. "
+            "-1 (default) auto-calculates as world_size / replicate_size."
+        ),
+    )
+    parser.add_argument(
+        "--hsdp-replicate-size",
+        type=int,
+        default=1,
+        help=(
+            "Number of replica groups for HSDP. Each replica holds a full sharded copy. "
+            "Default 1 means pure sharding (no replication)."
+        ),
+    )
     parser.add_argument(
         "--enforce-eager",
         action="store_true",
@@ -234,6 +257,9 @@ def main():
         vae_patch_parallel_size=args.vae_patch_parallel_size,
         ring_degree=args.ring_degree,
         ulysses_degree=args.ulysses_degree,
+        use_hsdp=args.use_hsdp,
+        hsdp_shard_size=args.hsdp_shard_size,
+        hsdp_replicate_size=args.hsdp_replicate_size,
     )
 
     profiler_enabled = args.profiler_config is not None

examples/offline_inference/text_to_audio/README.md

@@ -1,5 +1,12 @@
 # Text-To-Audio
 
+A unified script for text/video-to-audio generation. Supported models:
+
+| Model | Tasks | Notes |
+|-------|-------|-------|
+| `stabilityai/stable-audio-open-1.0` | text-to-audio | gated; uses `--audio-length` |
+| `zhangj1an/AudioX` | `t2a` / `t2m` / `v2a` / `v2m` / `tv2a` / `tv2m` | pass `--task`; video tasks need `--video` |
+
 The `stabilityai/stable-audio-open-1.0` pipeline generates audio from text prompts.
 
 ## Prerequisites
@@ -43,13 +50,45 @@ python text_to_audio.py \
   --output stable_audio_output.wav
 ```
 
+### AudioX
+
+AudioX supports six tasks. Sampler and reference knobs (declared in
+`vllm_omni/model_extras/audiox.py`) are passed via the generic `--extra-body`
+JSON flag, routed into sampling `extra_args`.
+
+Text tasks (`t2a` / `t2m`):
+
+```bash
+python text_to_audio.py \
+  --model zhangj1an/AudioX --task t2a \
+  --prompt "Fireworks burst twice, followed by a clock ticking." \
+  --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0 --seed 42 \
+  --extra-body '{"sigma_min": 0.03, "sigma_max": 1000.0}' \
+  --output t2a.wav
+```
+
+Video-conditioned tasks (`v2a` / `v2m` / `tv2a` / `tv2m`) require `--video`:
+
+```bash
+python text_to_audio.py \
+  --model zhangj1an/AudioX --task tv2a \
+  --prompt "drum beating sound and human talking" \
+  --video https://zeyuet.github.io/AudioX/static/samples/V2M/1XeBotOFqHA.mp4 \
+  --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0 \
+  --output tv2a.wav
+```
+
 Key arguments:
 
 - `--prompt`: text description (string).
+- `--task`: [AudioX] one of `t2a`/`t2m`/`v2a`/`v2m`/`tv2a`/`tv2m`.
+- `--video`: [AudioX `v2*`/`tv2*`] video file/URL for conditioning (→ `video_path`).
+- `--audio-start`: audio start offset in seconds (→ `audio_start_in_s` for Stable Audio, `seconds_start` for AudioX).
+- `--audio-length`: audio duration in seconds (audio length for Stable Audio, `seconds_total` for AudioX).
+- `--extra-body`: JSON dict of model-specific knobs (declared in `vllm_omni/model_extras/`), merged into sampling `extra_args`. For AudioX, sampler/reference knobs go here, e.g. `'{"sigma_min": 0.03, "sigma_max": 1000.0, "cfg_rescale": 0.0, "audio_path": "ref.wav"}'`.
 - `--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).
 - `--use-hsdp`: enable HSDP weight sharding for the Stable Audio DiT.
 - `--hsdp-shard-size`: number of GPUs used for HSDP sharding.

examples/offline_inference/text_to_audio/text_to_audio.py

@@ -2,37 +2,70 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 """
-Example script for text-to-audio generation using Stable Audio Open.
+Example script for text/video-to-audio generation with diffusion audio models.
 
-This script demonstrates how to generate audio from text prompts using
-the Stable Audio Open model with vLLM-Omni.
+This script supports:
+  * Stable Audio Open (`stabilityai/stable-audio-open-1.0`) — text-to-audio.
+  * AudioX (`zhangj1an/AudioX`) — six tasks: t2a / t2m / v2a / v2m / tv2a / tv2m.
+
+Model-specific generation knobs are declared in `vllm_omni/model_extras` and
+routed into `sampling_params.extra_args` via `apply_declared_extra_args`, so
+the same script serves both models without per-model branching in the engine.
 
 Usage:
+    # Stable Audio Open
     python text_to_audio.py --prompt "The sound of a dog barking"
     python text_to_audio.py --prompt "A piano playing a gentle melody" --audio-length 10.0
     python text_to_audio.py --prompt "Thunder and rain sounds" --negative-prompt "Low quality"
     python text_to_audio.py --prompt "A soft synth pad" --cache-backend tea_cache
+
+    # AudioX (text-to-audio / text-to-music)
+    python text_to_audio.py --model zhangj1an/AudioX --task t2a \
+        --prompt "Fireworks burst twice, then a clock ticking." \
+        --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0
+    # AudioX video-conditioned (v2*/tv2*) — pass a video file/URL
+    python text_to_audio.py --model zhangj1an/AudioX --task tv2a \
+        --prompt "drum beating sound and human talking" --video clip.mp4 \
+        --extra-body '{"sigma_min": 0.03, "sigma_max": 1000.0}'
 """
 
 import argparse
+import json
 import time
 from pathlib import Path
+from typing import Any
 
 import numpy as np
 import torch
 
 from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.diffusion.utils.param_utils import apply_declared_extra_args
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.model_extras import get_extra_body_params, get_model_class_name
 from vllm_omni.platforms import current_omni_platform
 
+AUDIOX_TASKS = ("t2a", "t2m", "v2a", "v2m", "tv2a", "tv2m")
+AUDIOX_VIDEO_TASKS = frozenset({"v2a", "v2m", "tv2a", "tv2m"})
+
+
+def parse_extra_body(value: str) -> dict[str, Any]:
+    """Parse a JSON object string for --extra-body."""
+    try:
+        obj = json.loads(value)
+    except json.JSONDecodeError as e:
+        raise argparse.ArgumentTypeError(f"--extra-body must be valid JSON: {e}") from e
+    if not isinstance(obj, dict):
+        raise argparse.ArgumentTypeError("--extra-body must be a JSON object")
+    return obj
+
 
 def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Generate audio with Stable Audio Open.")
+    parser = argparse.ArgumentParser(description="Generate audio with supported diffusion audio models.")
     parser.add_argument(
         "--model",
         default="stabilityai/stable-audio-open-1.0",
-        help="Stable Audio model name or local path.",
+        help="Audio diffusion model name or local path. Supported: stabilityai/stable-audio-open-1.0, zhangj1an/AudioX.",
     )
     parser.add_argument(
         "--prompt",
@@ -41,8 +74,11 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument(
         "--negative-prompt",
-        default="Low quality.",
-        help="Negative prompt for classifier-free guidance.",
+        default=None,
+        help=(
+            "Negative prompt for classifier-free guidance. Off by default; "
+            'recommended for Stable Audio, e.g. --negative-prompt "Low quality".'
+        ),
     )
     parser.add_argument(
         "--seed",
@@ -60,13 +96,19 @@ def parse_args() -> argparse.Namespace:
         "--audio-start",
         type=float,
         default=0.0,
-        help="Audio start time in seconds.",
+        help=(
+            "Audio start offset in seconds. Applies to both models "
+            "(audio_start_in_s for Stable Audio, seconds_start for AudioX)."
+        ),
     )
     parser.add_argument(
         "--audio-length",
         type=float,
         default=10.0,
-        help="Audio length in seconds (max ~47s for stable-audio-open-1.0).",
+        help=(
+            "Audio duration in seconds. Maps to audio length for Stable Audio "
+            "(max ~47s for stable-audio-open-1.0) and seconds_total for AudioX."
+        ),
     )
     parser.add_argument(
         "--num-inference-steps",
@@ -80,11 +122,41 @@ def parse_args() -> argparse.Namespace:
         default=1,
         help="Number of audio waveforms to generate for the given prompt.",
     )
+    parser.add_argument(
+        "--task",
+        type=str,
+        default=None,
+        choices=list(AUDIOX_TASKS),
+        help=(
+            "[AudioX only] Generation task: t2a/t2m/v2a/v2m/tv2a/tv2m. "
+            "v2*/tv2* require --video. Ignored for Stable Audio Open."
+        ),
+    )
+    parser.add_argument(
+        "--video",
+        type=str,
+        default=None,
+        help="[AudioX v2*/tv2*] Video file path or URL used for conditioning.",
+    )
+    parser.add_argument(
+        "--extra-body",
+        type=parse_extra_body,
+        default=None,
+        help=(
+            "Model-specific generation params as a JSON object, e.g. "
+            '\'{"sigma_min": 0.03, "sigma_max": 1000.0}\'. Each key is filtered '
+            "against the model's declared extra_body_params (see vllm_omni/model_extras), "
+            "so unknown keys are silently dropped. Values here override the equivalent flags."
+        ),
+    )
     parser.add_argument(
         "--output",
         type=str,
-        default="stable_audio_output.wav",
-        help="Path to save the generated audio (WAV format).",
+        default=None,
+        help=(
+            "Path to save the generated audio (WAV format). Defaults to "
+            "audiox_output.wav for AudioX and stable_audio_output.wav otherwise."
+        ),
     )
     parser.add_argument(
         "--sample-rate",
@@ -203,6 +275,8 @@ def save_audio(audio_data: np.ndarray, output_path: str, sample_rate: int = 4410
 
 def main():
     args = parse_args()
+    model_name = str(args.model).lower() if args.model else ""
+    is_audiox = "audiox" in model_name
     generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(args.seed)
     cache_config = None
     if args.cache_backend == "tea_cache":
@@ -211,12 +285,14 @@ def main():
         }
 
     print(f"\n{'=' * 60}")
-    print("Stable Audio Open - Text-to-Audio Generation")
+    print("Text/Video-to-Audio Generation")
     print(f"{'=' * 60}")
     print(f"  Model: {args.model}")
+    if args.task is not None:
+        print(f"  Task: {args.task}")
     print(f"  Prompt: {args.prompt}")
     print(f"  Negative prompt: {args.negative_prompt}")
-    print(f"  Audio length: {args.audio_length}s")
+    print(f"  Duration: {args.audio_length}s")
     print(f"  Inference steps: {args.num_inference_steps}")
     print(f"  Guidance scale: {args.guidance_scale}")
     print(f"  Cache backend: {args.cache_backend if args.cache_backend else 'None (no acceleration)'}")
@@ -235,8 +311,9 @@ def main():
         hsdp_replicate_size=args.hsdp_replicate_size,
     )
 
-    # Initialize Omni with Stable Audio model
-    omni = Omni(
+    # Initialize Omni. AudioX requires an explicit pipeline class; Stable Audio
+    # is resolved from its config.
+    omni_kwargs: dict[str, Any] = dict(
         model=args.model,
         parallel_config=parallel_config,
         cache_backend=args.cache_backend,
@@ -245,43 +322,68 @@ def main():
         enable_cpu_offload=args.enable_cpu_offload,
         enable_layerwise_offload=args.enable_layerwise_offload,
     )
+    if is_audiox:
+        omni_kwargs["model_class_name"] = "AudioXPipeline"
+    omni = Omni(**omni_kwargs)
+    model_class_name = get_model_class_name(omni)
+    declared_extra_body_params = get_extra_body_params(model_class_name)
 
-    # Calculate audio end time
-    audio_end_in_s = args.audio_start + args.audio_length
+    diffusion_params = OmniDiffusionSamplingParams(
+        generator=generator,
+        guidance_scale=args.guidance_scale,
+        num_inference_steps=args.num_inference_steps,
+        num_outputs_per_prompt=args.num_waveforms,
+        seed=args.seed,
+    )
+
+    # Negative prompt is opt-in for both models (off by default). Including a
+    # non-empty negative prompt enables the separate CFG conditioning branch.
+    prompt: dict[str, Any] = {"prompt": args.prompt}
+    if args.negative_prompt:
+        prompt["negative_prompt"] = args.negative_prompt
+
+    if model_class_name == "AudioXPipeline":
+        if args.task is None:
+            raise ValueError("AudioX requires --task (one of t2a/t2m/v2a/v2m/tv2a/tv2m).")
+        if args.task in AUDIOX_VIDEO_TASKS and not args.video:
+            raise ValueError(f"AudioX task {args.task!r} requires --video.")
+        # AudioX core knobs are surfaced as shared flags; sampler/reference knobs
+        # (sigma_min/sigma_max/cfg_rescale/audio_path) arrive via --extra-body and
+        # are filtered against the declared extra_body contract.
+        user_extra: dict[str, Any] = {
+            "audiox_task": args.task,
+            "seconds_start": args.audio_start,
+            "seconds_total": args.audio_length,
+            "video_path": args.video,
+        }
+        if args.extra_body:
+            user_extra.update(args.extra_body)
+        apply_declared_extra_args(diffusion_params, declared_extra_body_params, user_extra)
+    else:
+        # Stable Audio Open path (backward-compatible).
+        audio_end_in_s = args.audio_start + args.audio_length
+        diffusion_params.extra_args = {
+            "audio_start_in_s": args.audio_start,
+            "audio_end_in_s": audio_end_in_s,
+        }
+        if args.extra_body:
+            diffusion_params.extra_args.update(args.extra_body)
 
-    # Time profiling for generation
     generation_start = time.perf_counter()
 
-    # Generate audio
-    outputs = omni.generate(
-        {
-            "prompt": args.prompt,
-            "negative_prompt": args.negative_prompt,
-        },
-        OmniDiffusionSamplingParams(
-            generator=generator,
-            guidance_scale=args.guidance_scale,
-            num_inference_steps=args.num_inference_steps,
-            num_outputs_per_prompt=args.num_waveforms,
-            extra_args={
-                "audio_start_in_s": args.audio_start,
-                "audio_end_in_s": audio_end_in_s,
-            },
-        ),
-    )
+    outputs = omni.generate(prompt, diffusion_params)
 
     generation_end = time.perf_counter()
     generation_time = generation_end - generation_start
 
     print(f"Total generation time: {generation_time:.2f} seconds")
 
-    # Process and save audio
-    output_path = Path(args.output)
+    output = args.output or ("audiox_output.wav" if model_class_name == "AudioXPipeline" else "stable_audio_output.wav")
+    output_path = Path(output)
     output_path.parent.mkdir(parents=True, exist_ok=True)
     suffix = output_path.suffix or ".wav"
     stem = output_path.stem or "stable_audio_output"
 
-    # Extract audio from omni.generate() outputs
     if not outputs:
         raise ValueError("No output generated from omni.generate()")
 
@@ -296,7 +398,6 @@ def main():
     if audio is None:
         raise ValueError("No audio output found in request_output")
 
-    # Handle different output formats
     if isinstance(audio, torch.Tensor):
         audio = audio.cpu().float().numpy()
 
@@ -323,7 +424,7 @@ def main():
         save_audio(audio, str(output_path), args.sample_rate)
         print(f"Saved generated audio to {output_path}")
 
-    print(f"\nGenerated {args.audio_length}s of audio at {args.sample_rate} Hz")
+    print(f"\nGenerated ~{args.audio_length}s of audio at {args.sample_rate} Hz")
 
 
 if __name__ == "__main__":

examples/offline_inference/text_to_image/README.md

@@ -179,7 +179,6 @@ python examples/offline_inference/text_to_image/text_to_image.py \
   --output flux2-dev.png
 ```
 
-
 ### HiDream-I1-Full Models
 
 The `--auxiliary-text-encoder` parameter is required when running HiDream‑I1‑Full:
@@ -199,7 +198,9 @@ python examples/offline_inference/text_to_image/text_to_image.py \
 
 ### Batch Requests (Multiple Prompts)
 
-You can pass multiple prompts in a single `generate` call.
+You can pass multiple prompts in a single `Omni.generate` call. `Omni`
+submits each prompt as an independent request and returns one output per
+prompt.
 
 ```python
 from vllm_omni.entrypoints.omni import Omni
@@ -224,11 +225,10 @@ if __name__ == "__main__":
 
 !!! info
 
-    For diffusion pipelines, the stage config field `stage_args.[].runtime.max_batch_size` is 1 by
-    default, and the input list is sliced into single-item requests before feeding into the diffusion
-    pipeline. For models that do internally support batched inputs, you can
-    [modify this configuration](../../../configuration/stage_configs.md) to let the model accept a
-    longer batch of prompts.
+    For diffusion pipelines, the input list is sliced into single-item requests
+    before feeding into the diffusion pipeline. For request-level batching
+    controls such as `max_num_seqs`, see
+    [Request-Level Batching](../../../docs/user_guide/diffusion/request_batching.md).
 
 ### Negative Prompts
 

examples/offline_inference/text_to_image/text_to_image.py

@@ -69,7 +69,16 @@ def parse_args() -> argparse.Namespace:
         "--stage-configs-path",
         type=str,
         default=None,
-        help="Path to a YAML file containing stage configurations for Omni.",
+        help="[Deprecated] Path to a legacy stage_args-format YAML. Prefer --deploy-config.",
+    )
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help=(
+            "Path to a deploy YAML (new pipeline/stages format). Required for multi-stage "
+            "text-to-image pipelines whose deploy config is not auto-loaded."
+        ),
     )
     parser.add_argument("--prompt", default="a cup of coffee on the table", help="Text prompt for image generation.")
     parser.add_argument(
@@ -155,7 +164,12 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--enforce-eager",
         action="store_true",
-        help="Disable torch.compile and force eager execution.",
+        default=None,
+        help=(
+            "Disable torch.compile and force eager execution. Left unset (None) "
+            "so it is only forwarded when explicitly given; "
+            "otherwise the per-stage deploy YAML value wins."
+        ),
     )
     parser.add_argument(
         "--enable-cpu-offload",
@@ -221,8 +235,13 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--tensor-parallel-size",
         type=int,
-        default=1,
-        help="Number of GPUs used for tensor parallelism (TP) inside the DiT.",
+        default=None,
+        help=(
+            "Number of GPUs used for tensor parallelism (TP) inside the DiT. "
+            "Left unset so it is only forwarded when explicitly given; "
+            "otherwise the per-stage deploy YAML (or engine default of 1) wins. "
+            "Passing it always overrides the deploy YAML."
+        ),
     )
     parser.add_argument(
         "--enable-expert-parallel",
@@ -410,10 +429,8 @@ def main():
         "ring_degree": args.ring_degree,
         "ulysses_mode": args.ulysses_mode,
         "cfg_parallel_size": args.cfg_parallel_size,
-        "tensor_parallel_size": args.tensor_parallel_size,
         "vae_patch_parallel_size": args.vae_patch_parallel_size,
         "enable_expert_parallel": args.enable_expert_parallel,
-        "enforce_eager": args.enforce_eager,
         "enable_cpu_offload": args.enable_cpu_offload,
         "mode": "text-to-image",
         "log_stats": args.log_stats,
@@ -425,8 +442,14 @@ def main():
         **lora_args,
         **quant_kwargs,
     }
+    if args.tensor_parallel_size is not None:
+        omni_kwargs["tensor_parallel_size"] = args.tensor_parallel_size
+    if args.enforce_eager is not None:
+        omni_kwargs["enforce_eager"] = args.enforce_eager
     if args.stage_configs_path:
         omni_kwargs["stage_configs_path"] = args.stage_configs_path
+    if args.deploy_config:
+        omni_kwargs["deploy_config"] = args.deploy_config
     if use_nextstep:
         # NextStep-1.1 requires explicit pipeline class
         omni_kwargs["model_class_name"] = "NextStep11Pipeline"
@@ -451,8 +474,9 @@ def main():
     print(f"  Quantization: {args.quantization if args.quantization else 'None (BF16)'}")
     if ignored_layers:
         print(f"  Ignored layers: {ignored_layers}")
+    tp_display = args.tensor_parallel_size if args.tensor_parallel_size is not None else "deploy/default"
     print(
-        f"  Parallel configuration: tensor_parallel_size={args.tensor_parallel_size}, "
+        f"  Parallel configuration: tensor_parallel_size={tp_display}, "
         f"ulysses_degree={args.ulysses_degree}, ulysses_mode={args.ulysses_mode}, "
         f"ring_degree={args.ring_degree}, cfg_parallel_size={args.cfg_parallel_size}, "
         f"vae_patch_parallel_size={args.vae_patch_parallel_size}, "
@@ -522,8 +546,8 @@ def main():
         diffusion_params.extra_args.update({k: v for k, v in user_extra.items() if v is not None})
 
     if lora_request:
-        diffusion_params.extra_args["lora_request"] = lora_request
-        diffusion_params.extra_args["lora_scale"] = args.lora_scale
+        diffusion_params.lora_request = lora_request
+        diffusion_params.lora_scale = args.lora_scale
 
     # Build per-stage sampling params for multi-stage models (e.g. BAGEL),
     # or wrap single diffusion params for single-stage models.
@@ -589,6 +613,12 @@ def main():
         if images:
             break
 
+    # Fallback: generation-stage pipelines (e.g. MammothModa2's AR->DiT) return the
+    # generated image as a tensor under multimodal_output instead of populating the
+    # `images` field that diffusion-stage pipelines fill.
+    if not images:
+        images = _images_from_multimodal_output(outputs)
+
     if not images:
         raise ValueError("No images found in request_output")
 
@@ -606,5 +636,36 @@ def main():
             print(f"Saved generated image to {save_path}")
 
 
+def _images_from_multimodal_output(outputs: list[Any]) -> list[Any]:
+    """Extract PIL images from multimodal_output tensors.
+
+    Generation-stage pipelines (e.g. MammothModa2's AR->DiT) return the generated
+    image as a tensor (normalized to [-1, 1], CHW) under ``multimodal_output``
+    rather than populating the ``images`` field. Convert any such tensors to PIL.
+    """
+    from PIL import Image
+
+    pil_images: list[Any] = []
+    for output in outputs:
+        req_out = getattr(output, "request_output", output)
+        for completion in getattr(req_out, "outputs", None) or []:
+            # multimodal_output is a MultimodalPayload (a Mapping) keyed by modality,
+            # matching how omni examples (ming_flash_omni / magi_human / dynin) read it.
+            mm = getattr(completion, "multimodal_output", None) or {}
+            if "image" not in mm:
+                continue
+            payload = mm["image"]
+            for tensor in payload if isinstance(payload, list) else [payload]:
+                if not isinstance(tensor, torch.Tensor):
+                    continue
+                img = tensor.detach().to("cpu", dtype=torch.float32)
+                if img.ndim == 4:
+                    img = img[0]
+                img = (img / 2 + 0.5).clamp(0, 1).mul(255).to(torch.uint8)
+                img = img.permute(1, 2, 0).contiguous().numpy()
+                pil_images.append(Image.fromarray(img))
+    return pil_images
+
+
 if __name__ == "__main__":
     main()

examples/offline_inference/text_to_speech/glm_tts/end2end.py

@@ -101,7 +101,7 @@ def _extract_sample_rate(audio_mm: dict) -> int:
 def _modify_deploy_config(base_path: str, async_chunk: bool) -> str:
     """Build deploy config with explicit sync/async mode and eager execution.
 
-    Mirrors the logic in ``tests/e2e/offline_inference/test_glm_tts.py``
+    Mirrors the logic in ``tests/e2e/offline_inference/test_glm_tts_expansion.py``
     (``_get_deploy_config``) so that example runs match CI behavior.
     """
     with open(base_path) as f:

examples/offline_inference/text_to_speech/omnivoice/end2end.py

@@ -34,7 +34,7 @@ def run_e2e():
         help="Model name or path (HuggingFace or local)",
     )
     parser.add_argument(
-        "--stage-config",
+        "--deploy-config",
         type=str,
         default="vllm_omni/deploy/omnivoice.yaml",
     )
@@ -87,14 +87,14 @@ def run_e2e():
     )
     args = parser.parse_args()
 
-    if not os.path.exists(args.stage_config):
-        raise FileNotFoundError(f"Stage config not found: {args.stage_config}")
+    if not os.path.exists(args.deploy_config):
+        raise FileNotFoundError(f"Deploy config not found: {args.deploy_config}")
 
     print(f"Initializing OmniVoice with model={args.model}")
 
     omni = Omni(
         model=args.model,
-        stage_configs_path=args.stage_config,
+        deploy_config=args.deploy_config,
         log_stats=True,
     )
 

examples/offline_inference/text_to_video/text_to_video.md

@@ -8,6 +8,7 @@ A unified script for text-to-video generation. Supports multiple models with mod
 |---|---|---|---|---|---|
 | `Wan-AI/Wan2.1-VACE-1.3B-diffusers` | 480x832 | 81 | 30 | 5.0 | ~20 GiB (RTX 5090, VAE tiling) |
 | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 720x1280 | 81 | 40 | 4.0 | ~60 GiB |
+| `dg845/LTX-2.3-Diffusers` | 384x512 | 25 | 20 | 4.0 | 96GB-class GPU |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | 480x832 | 121 | 50 | 6.0 | 1×A100 80GB |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
 | `nvidia/Cosmos3-Nano` | 720x1280 | 189 | 35 | 6.0 | ~46 GiB (peak, 720p) |
@@ -70,6 +71,30 @@ python text_to_video.py \
   --output ltx2_out.mp4
 ```
 
+### LTX-2.3
+
+```bash
+python text_to_video.py \
+  --model dg845/LTX-2.3-Diffusers \
+  --model-class-name LTX23Pipeline \
+  --prompt "Cherry blossoms swaying gently in the breeze with synchronized ambient sound" \
+  --negative-prompt "worst quality, inconsistent motion, blurry, jittery, distorted" \
+  --height 384 \
+  --width 512 \
+  --num-frames 25 \
+  --num-inference-steps 20 \
+  --guidance-scale 4.0 \
+  --frame-rate 24 \
+  --fps 24 \
+  --audio-sample-rate 48000 \
+  --output ltx23_t2v_output.mp4
+```
+
+Use a Diffusers-format checkpoint such as `dg845/LTX-2.3-Diffusers`; the
+upstream `Lightricks/LTX-2.3` raw safetensors repo is not directly loadable by
+this pipeline. Pass `--model-class-name LTX23Pipeline` to select the LTX-2.3
+text-to-video pipeline explicitly.
+
 ### HunyuanVideo-1.5 (480p)
 
 ```bash
@@ -187,6 +212,8 @@ python text_to_video.py \
 ### Common
 
 - `--model`: Diffusers model ID or local path.
+- `--model-class-name`: Optional explicit pipeline class. Use `LTX23Pipeline`
+  for LTX-2.3 text-to-video.
 - `--prompt`: text description (string).
 - `--height/--width`: output resolution. Default depends on model.
 - `--num-frames`: number of frames. Default depends on model.
@@ -202,7 +229,8 @@ python text_to_video.py \
 - `--enable-cpu-offload`: enable CPU offloading for diffusion models.
 - `--enable-layerwise-offload`: enable layerwise offloading on DiT modules.
 - `--frame-rate`: generation FPS for pipelines that require it (e.g., LTX2).
-- `--audio-sample-rate`: audio sample rate for embedded audio (when the pipeline returns audio).
+- `--audio-sample-rate`: audio sample rate for embedded audio (when the
+  pipeline returns audio; LTX-2.3 outputs 48kHz audio).
 - `--quantization`: quantization method (`fp8` for FP8, `gguf` for GGUF).
 - `--flow-shift`: scheduler flow_shift parameter.
 - `--extra-body`: JSON object of model-specific generation params, filtered against the model's declared `extra_body_params` (see [`vllm_omni/model_extras`](../../../vllm_omni/model_extras)). Used by Cosmos3 (see above).

examples/offline_inference/text_to_video/text_to_video.py

@@ -455,6 +455,7 @@ def main():
         print(f"Worker peak GPU memory (reserved): {peak_mb:.2f} MiB ({peak_mb / 1024:.2f} GiB)")
 
     audio = None
+    audio_sample_rate = args.audio_sample_rate
     if isinstance(frames, list):
         frames = frames[0] if frames else None
 
@@ -465,11 +466,13 @@ def main():
             )
         if frames.multimodal_output and "audio" in frames.multimodal_output:
             audio = frames.multimodal_output["audio"]
+            audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
         if frames.is_pipeline_output and frames.request_output is not None:
             inner_output = frames.request_output
             if isinstance(inner_output, OmniRequestOutput):
                 if inner_output.multimodal_output and "audio" in inner_output.multimodal_output:
                     audio = inner_output.multimodal_output["audio"]
+                    audio_sample_rate = inner_output.multimodal_output.get("audio_sample_rate", audio_sample_rate)
                 frames = inner_output
         if isinstance(frames, OmniRequestOutput):
             if frames.images:
@@ -477,6 +480,7 @@ def main():
                     frames, audio = frames.images[0]
                 elif len(frames.images) == 1 and isinstance(frames.images[0], dict):
                     audio = frames.images[0].get("audio")
+                    audio_sample_rate = frames.images[0].get("audio_sample_rate", audio_sample_rate)
                     frames = frames.images[0].get("frames") or frames.images[0].get("video")
                 else:
                     frames = frames.images
@@ -489,6 +493,7 @@ def main():
             frames, audio = first_item
         elif isinstance(first_item, dict):
             audio = first_item.get("audio")
+            audio_sample_rate = first_item.get("audio_sample_rate", audio_sample_rate)
             frames = first_item.get("frames") or first_item.get("video")
         elif isinstance(first_item, list):
             frames = first_item
@@ -497,6 +502,7 @@ def main():
         frames, audio = frames
     elif isinstance(frames, dict):
         audio = frames.get("audio")
+        audio_sample_rate = frames.get("audio_sample_rate", audio_sample_rate)
         frames = frames.get("frames") or frames.get("video")
 
     if frames is None:
@@ -611,7 +617,7 @@ def main():
             frames_u8,
             audio_np,
             fps=float(args.fps),
-            audio_sample_rate=args.audio_sample_rate,
+            audio_sample_rate=audio_sample_rate,
         )
         with open(str(output_path), "wb") as f:
             f.write(video_bytes)

examples/online_serving/audiox/README.md

@@ -37,11 +37,10 @@ python openai_chat_client.py --task tv2a \
 
 The client sends:
 
-- `num_inference_steps`, `guidance_scale`, `seed` as first-class OpenAI chat-completion fields
-- `audiox_task`, `seconds_start`, `seconds_total`, `sigma_min`, `sigma_max` nested under
-  `extra_args` (a reserved dict on the request body that the server forwards verbatim into
-  the pipeline's `sampling_params.extra_args` — the same escape hatch `serving_video.py` exposes
-  as `extra_params` on /v1/videos)
+- `num_inference_steps`, `guidance_scale`, `seed`, `audiox_task`, `seconds_start`,
+  `seconds_total`, `sigma_min`, `sigma_max` as flattened keys under `extra_body`. The server
+  routes the declared knobs (see `vllm_omni/model_extras/audiox.py`) into the pipeline's
+  `sampling_params.extra_args`.
 - For `v2*` / `tv2*` tasks, the video as a `video_url` content item (data URI for local files)
 
 ## curl
@@ -52,10 +51,10 @@ curl -sS -X POST http://localhost:8099/v1/chat/completions \
   -d '{
     "model": "zhangj1an/AudioX",
     "messages": [{"role": "user", "content": [{"type": "text", "text": "Uplifting ukulele"}]}],
-    "num_inference_steps": 250,
-    "guidance_scale": 7.0,
-    "seed": 42,
-    "extra_args": {
+    "extra_body": {
+      "num_inference_steps": 250,
+      "guidance_scale": 7.0,
+      "seed": 42,
       "audiox_task": "t2m",
       "seconds_total": 10.0,
       "sigma_min": 0.3,

examples/online_serving/audiox/openai_chat_client.py

@@ -4,8 +4,8 @@
 AudioX supports 6 tasks (t2a, t2m, v2a, v2m, tv2a, tv2m). Text-only tasks send the prompt as the
 chat message; video-conditioned tasks additionally attach the video via a ``video_url`` content
 item (data URI for local files). Task + generation knobs (steps, cfg, sigma range, seconds, seed)
-are sent via the OpenAI SDK's ``extra_body`` as ``extra_args`` — the same pipeline-agnostic escape
-hatch used by the /v1/videos endpoint's ``extra_params`` field.
+are sent as flattened keys under ``extra_body``; the server routes the declared knobs into
+``sampling_params.extra_args`` (see ``vllm_omni/model_extras/audiox.py``).
 
 Usage:
   python openai_chat_client.py --task t2a --prompt "Fireworks burst twice..." --output t2a.wav
@@ -88,10 +88,10 @@ def main() -> int:
     payload = {
         "model": args.model,
         "messages": [{"role": "user", "content": content}],
-        "num_inference_steps": args.steps,
-        "guidance_scale": args.guidance_scale,
-        "seed": args.seed,
-        "extra_args": {
+        "extra_body": {
+            "num_inference_steps": args.steps,
+            "guidance_scale": args.guidance_scale,
+            "seed": args.seed,
             "audiox_task": args.task,
             "seconds_start": args.seconds_start,
             "seconds_total": args.seconds_total,

examples/online_serving/diffusers_pipeline_adapter/README.md

@@ -6,7 +6,8 @@ vLLM-Omni supports running diffusion models with the diffusers backend, directly
 
 The diffusers backend is a black-box adapter. Its primary focus is to serve diffusion models online.
 Currently, the following features are NOT yet supported.
-It is not guaranteed whether they will be supported in the future.
+Community contributions are welcome to add these features when they can be
+implemented with clear Diffusers or vLLM-Omni behavior.
 
 - CFG parallel execution
 - Sequence parallel execution
@@ -68,6 +69,35 @@ But if that parameter is simultaneously set in both the vLLM-Omni interface and
     Consider referring to [`OmniDiffusionSamplingParams`](../../../vllm_omni/inputs/data.py) for the default sampling parameters in the vLLM-Omni interface,
     and the corresponding [diffusers pipeline](https://huggingface.co/docs/diffusers/main/en/api/pipelines/overview)'s `__call__` function documentation.
 
+### Quantization
+
+Use `diffusers_load_kwargs` for Diffusers-native quantization options. If
+`diffusers_load_kwargs["quantization_config"]` is provided as a dictionary, the
+diffusers backend builds a Diffusers quantization config and lets Diffusers
+validate it before calling `DiffusionPipeline.from_pretrained()`.
+
+The diffusers backend also provides a small compatibility shortcut for
+vLLM-Omni quantization configs when `diffusers_load_kwargs` does not already
+contain `quantization_config`. Currently this maps online/dynamic `fp8` and
+online/dynamic `int8` to Diffusers/TorchAO dynamic quantization for transformer
+components such as `transformer` or `transformer_2`. This path requires
+`torchao` to be installed.
+
+For example, the CLI can request dynamic FP8 through the vLLM-Omni interface:
+
+```bash
+vllm serve "Qwen/Qwen-Image" \
+    --omni \
+    --diffusion-load-format diffusers \
+    --quantization-config '{"method": "fp8"}'
+```
+
+Other vLLM-Omni quantization methods, such as `gguf`, `modelopt`, `mxfp4`,
+`mxfp8`, serialized checkpoints, static FP8 configs, and layer-name skip lists
+such as `ignored_layers`, are intentionally not translated by this compatibility
+shortcut. Use Diffusers-native configuration through `diffusers_load_kwargs` or
+a native vLLM-Omni pipeline for those cases.
+
 ### Attention Backends
 
 The diffusers backend converts

examples/online_serving/ming_flash_omni/README.md

@@ -49,22 +49,14 @@ python examples/online_serving/openai_chat_completion_client_for_multimodal_gene
     --modalities text
 ```
 
-Parameterized curl wrapper in this directory:
 
-```bash
-bash run_curl_multimodal_generation.sh text
-bash run_curl_multimodal_generation.sh use_image
-bash run_curl_multimodal_generation.sh use_audio
-bash run_curl_multimodal_generation.sh use_video
-bash run_curl_multimodal_generation.sh use_mixed_modalities
-bash run_curl_multimodal_generation.sh use_image_gen
-```
+## Image generation (text-to-image / img2img)
 
-## Image generation (text-to-image)
+Ming-flash-omni-2.0 also exposes an image-generation (diffusion) stage. Launch with the image deploy YAML, which adds an image-gen stage behind the thinker.
 
-Ming-flash-omni-2.0 also exposes an image-generation (diffusion) stage.
-Launch with the image deploy YAML, which adds an image-gen stage behind
-the thinker:
+The image-generation stage is a standard vLLM-Omni diffusion pipeline (`MingImagePipeline`); its request knobs are declared in `vllm_omni/model_extras/ming_flash_omni.py` and routed through `extra_body`, so they no longer need a bespoke `sampling_params_list` recipe (that form is still available for per-stage thinker sampling — see below).
+
+### Launch
 
 ```bash
 vllm serve Jonathan1909/Ming-flash-omni-2.0 --omni \
@@ -74,32 +66,58 @@ vllm serve Jonathan1909/Ming-flash-omni-2.0 --omni \
     --port 8091
 ```
 
-Then request image output by passing `"modalities": ["image"]`:
+
+### Text-to-image
+
+Request image output with `"modalities": ["image"]`:
 
 ```bash
 curl -s http://127.0.0.1:8091/v1/chat/completions \
-    -H "Content-Type: application/json" \
-    -d '{
-      "model": "Jonathan1909/Ming-flash-omni-2.0",
-      "messages": [{"role": "user", "content": "Please draw a cute cat."}],
-      "modalities": ["image"]
-    }' | jq -r '.choices[0].message.content[0].image_url.url | split(",")[1]' | base64 -d > ming_imagegen.png
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "Jonathan1909/Ming-flash-omni-2.0",
+    "messages": [
+      {
+        "role": "user",
+        "content": "Please draw a cute cat."
+      }
+    ],
+    "modalities": ["image"]
+  }' \
+  | jq -r '.choices[0].message.content[0].image_url.url | split(",")[1]' \
+  | base64 -d > ming_imagegen.png
 ```
 
-### Optional knobs
+Pass generation knobs under a literal `extra_body` object (the OpenAI Python client's `extra_body=` kwarg produces the same request). Keys are filtered against the declared set and routed into every stage's `extra_args`:
 
-Pass image-gen overrides as flat keys on the diffusion-stage `sampling_params_list[1].extra_args`:
+```bash
+curl http://127.0.0.1:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "Jonathan1909/Ming-flash-omni-2.0",
+    "modalities": ["image"],
+    "messages": [
+      {
+        "role": "user",
+        "content": "Draw a poster."
+      }
+    ],
+    "extra_body": {
+      "steps": 6,
+      "cfg": 1.5,
+      "height": 512,
+      "width": 512,
+      "seed": 123,
+      "byte5_text": ["理解与生成统一"],
+      "negative_prompt": "ugly, blurry, distorted"
+    }
+  }' \
+  | jq -r '.choices[0].message.content[0].image_url.url | split(",")[1]' \
+  | base64 -d > ming_imagegen_extra_body.png
+```
 
-| Key | Default | Description |
-| --- | --- | --- |
-| `height` / `width` | from config (1024) | Output resolution (multiples of `vae_scale_factor * 2`, currently 16). |
-| `steps` | 30 | Number of FlowMatchEuler denoise steps. |
-| `cfg` | 2.0 | Classifier-free guidance scale. |
-| `seed` | 42 | Per-request RNG seed (deterministic when ByT5 is also seed-stable). |
-| `byte5_text` | (auto) | Override the glyph text for ByT5 enhancement; raw strings are auto-wrapped to Ming's `Text "...". ` format. |
-| `negative_prompt` | empty | Real CFG negative conditioning (set on **stage-0 thinker** `extra_args` so `expand_cfg_prompts` spawns the companion). |
-
-Example with all the knobs:
+NOTE: `extra_body` does **not** set the thinker's own sampling params. To tune the thinker (stage-0) sampling (`temperature` / `top_p` / `top_k` / `max_tokens`) or place knobs explicitly per stage, use `sampling_params_list` (`[thinker, imagegen]`).
+`negative_prompt` must sit on the **stage-0** entry to trigger the real-CFG companion; the imagegen knobs go on **stage-1**:
 
 ```bash
 curl http://127.0.0.1:8091/v1/chat/completions \
@@ -143,27 +161,48 @@ curl http://127.0.0.1:8091/v1/chat/completions \
 
 ### img2img (reference image)
 
-Add an `image_url` content part to the user message; the parser routes it
-into the diffusion stage as `extra[reference_image]`:
+Add an `image_url` content part (a base64 data URL) to the user message; it is routed into the DiT stage as `extra[reference_image]`. base64-encode a local file and stream it through `jq` via stdin - piping `base64 -> jq -> curl` so that avoids the shell `ARG_MAX` limit that inlining a large base64 string in `-d '…'` hits.
 
-```jsonc
-"messages": [{
-  "role": "user",
-  "content": [
-    {"type": "text", "text": "Change the background to a sandy beach at sunset."},
-    {"type": "image_url", "image_url": {"url": "data:image/png;base64,<base64>"}}
-  ]
-}]
+```bash
+# Reference image: figures/cases/person_gen_05.png from the upstream Ming repo
+# Check https://github.com/inclusionAI/Ming/blob/3954fcb880ff5e61ff128bcf7f1ec344d46a6fe3/examples/vllm_demo.py
+wget https://raw.githubusercontent.com/inclusionAI/Ming/3954fcb880ff5e61ff128bcf7f1ec344d46a6fe3/figures/cases/person_gen_05.png
+
+base64 -w0 person_gen_05.png \
+| jq -Rs --arg prompt "Put a pair of sunglasses on the person." '{
+    model: "Jonathan1909/Ming-flash-omni-2.0",
+    modalities: ["image"],
+    messages: [
+      {
+        role: "user",
+        content: [
+          { type: "text", text: $prompt },
+          { type: "image_url", image_url: { url: ("data:image/png;base64," + (. | rtrimstr("\n"))) } }
+        ]
+      }
+    ]
+  }' \
+| curl http://127.0.0.1:8091/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d @- \
+| jq -r '.choices[0].message.content[0].image_url.url | split(",")[1]' \
+| base64 -d > ming_img2img.png
 ```
 
-### GPU layout
+The reference image can also be a public URL (`"url": "https://…/photo.jpg"`) or the simplified `{"image": "<base64>"}` content-part form — see the [image-to-image request formats](../image_to_image/README.md#request-format).
 
-The shipped `ming_flash_omni_image.yaml` allocates the thinker on GPUs 0–3
-(TP=4) and the diffusion stage on GPU 4 (TP=1). Copy the YAML and edit
-`devices` per stage to relocate; with fewer GPUs available, drop
-the thinker TP to 2 and run the diffusion stage on a free card. Image-gen
-warmup takes roughly an extra 30–60 s on top of the thinker — set
-`--stage-init-timeout 1800` if the default 300 s is too tight.
+### Knobs (declared `extra_body` params)
+
+| Key | Default | Description |
+| --- | --- | --- |
+| `height` / `width` | 1024 | Output resolution (multiples of `vae_scale_factor * 2`, currently 16). |
+| `steps` | 30 | Number of FlowMatchEuler denoise steps. |
+| `cfg` | 2.0 | Classifier-free guidance scale. |
+| `seed` | 42 | Per-request RNG seed. |
+| `byte5_text` | (auto) | Glyph text for ByT5 enhancement; raw strings are auto-wrapped to Ming's `Text "…". ` format. Auto-extracted from quoted spans in the prompt when omitted. |
+| `negative_prompt` | (empty) | Real CFG negative conditioning (text-to-image only). |
+
+For the offline `text_to_image.py` / `image_edit.py` scripts and the full knob reference, see the [image-generation section in the recipe](../../../recipes/inclusionAI/Ming-flash-omni-2.0.md#image-generation-text-to-image--img2img).
 
 ## Modality control
 

examples/online_serving/ming_flash_omni/run_curl_multimodal_generation.sh

@@ -1,169 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-# Server port
-PORT="${PORT:-8091}"
-# Default query type
-QUERY_TYPE="${1:-text}"
-
-# Validate query type
-if [[ ! "$QUERY_TYPE" =~ ^(text|use_audio|use_image|use_video|use_mixed_modalities|use_image_gen)$ ]]; then
-    echo "Error: Invalid query type '$QUERY_TYPE'"
-    echo "Usage: $0 [text|use_audio|use_image|use_video|use_mixed_modalities|use_image_gen]"
-    echo "  text: Text-only query"
-    echo "  use_audio: Audio + Text query"
-    echo "  use_image: Image + Text query"
-    echo "  use_video: Video + Text query"
-    echo "  use_mixed_modalities: Audio + Image + Video + Text query"
-    echo "  use_image_gen: Text-to-image (diffusion stage) — saves the generated PNG locally"
-    exit 1
-fi
-
-# Define URLs for assets
-MARY_HAD_LAMB_AUDIO_URL="https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/mary_had_lamb.ogg"
-CHERRY_BLOSSOM_IMAGE_URL="https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/cherry_blossom.jpg"
-SAMPLE_VIDEO_URL="https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/sample_demo_1.mp4"
-
-# Build user content based on query type
-case "$QUERY_TYPE" in
-  text)
-    user_content='[
-      {
-        "type": "text",
-        "text": "请详细介绍鹦鹉的生活习性。"
-      }
-    ]'
-    ;;
-  use_image)
-    user_content='[
-        {
-          "type": "image_url",
-          "image_url": {
-            "url": "'"$CHERRY_BLOSSOM_IMAGE_URL"'"
-          }
-        },
-        {
-          "type": "text",
-          "text": "Describe this image in detail."
-        }
-      ]'
-    ;;
-  use_audio)
-    user_content='[
-        {
-          "type": "audio_url",
-          "audio_url": {
-            "url": "'"$MARY_HAD_LAMB_AUDIO_URL"'"
-          }
-        },
-        {
-          "type": "text",
-          "text": "Please recognize the language of this speech and transcribe it. Format: oral."
-        }
-      ]'
-    ;;
-  use_video)
-    user_content='[
-        {
-          "type": "video_url",
-          "video_url": {
-            "url": "'"$SAMPLE_VIDEO_URL"'"
-          }
-        },
-        {
-          "type": "text",
-          "text": "Describe what is happening in this video."
-        }
-      ]'
-    ;;
-  use_mixed_modalities)
-    user_content='[
-        {
-          "type": "image_url",
-          "image_url": {
-            "url": "'"$CHERRY_BLOSSOM_IMAGE_URL"'"
-          }
-        },
-        {
-          "type": "audio_url",
-          "audio_url": {
-            "url": "'"$MARY_HAD_LAMB_AUDIO_URL"'"
-          }
-        },
-        {
-          "type": "text",
-          "text": "Describe the image, and recognize the language of this speech and transcribe it. Format: oral"
-        }
-      ]'
-    ;;
-esac
-
-echo "Running query type: $QUERY_TYPE"
-echo ""
-
-if [[ "$QUERY_TYPE" == "use_image_gen" ]]; then
-    # Image-gen branch: dual-stage diffusion endpoint, modalities=["image"].
-    # The shipped ming_flash_omni_image.yaml runs the thinker on cards 0-3
-    # (TP=4) and the diffusion stage on card 4 (TP=1). See the image-gen
-    # section in README.md for knobs and img2img.
-    out_path="${OUT_PATH:-/tmp/ming_imagegen.png}"
-    request_body=$(cat <<'EOF'
-{
-  "model": "Jonathan1909/Ming-flash-omni-2.0",
-  "modalities": ["image"],
-  "messages": [
-    {"role": "user", "content": "Draw a beautiful girl with short black hair and red dress."}
-  ]
-}
-EOF
-)
-    output=$(curl -sS --retry 3 --retry-delay 3 --retry-connrefused \
-        -X POST "http://localhost:${PORT}/v1/chat/completions" \
-        -H "Content-Type: application/json" \
-        -d "$request_body")
-    # Server returns base64-encoded PNG either as a data URL in
-    # ``choices[0].message.content[0].image_url.url`` or as a raw string.
-    b64=$(echo "$output" | jq -r '.choices[0].message.content
-        | if type == "array"
-            then (map(select(.image_url? != null)) | .[0].image_url.url // "")
-            elif type == "string" then . else "" end')
-    if [[ -z "$b64" ]]; then
-        echo "Error: no image returned. Raw response:"
-        echo "$output" | jq '.'
-        exit 1
-    fi
-    if [[ "$b64" == data:image* ]]; then b64="${b64#*,}"; fi
-    printf '%s' "$b64" | base64 --decode > "$out_path"
-    echo "Saved generated image: $out_path ($(stat -c%s "$out_path" 2>/dev/null || stat -f%z "$out_path") bytes)"
-    exit 0
-fi
-
-request_body=$(cat <<EOF
-{
-  "model": "Jonathan1909/Ming-flash-omni-2.0",
-  "modalities": ["text"],
-  "messages": [
-    {
-      "role": "system",
-      "content": [
-        {
-          "type": "text",
-          "text": "你是一个友好的AI助手。\n\ndetailed thinking off"
-        }
-      ]
-    },
-    {
-      "role": "user",
-      "content": $user_content
-    }
-  ]
-}
-EOF
-)
-
-output=$(curl -sS --retry 3 --retry-delay 3 --retry-connrefused \
-    -X POST http://localhost:${PORT}/v1/chat/completions \
-    -H "Content-Type: application/json" \
-    -d "$request_body")
-
-echo "Output of request: $(echo "$output" | jq '.choices[0].message.content')"

examples/online_serving/sensenova_u1/README.md

@@ -6,11 +6,11 @@
 vllm serve SenseNova/SenseNova-U1-8B-MoT --omni --port 8091
 ```
 
-With Cache-DiT acceleration:
+With cache acceleration (`cache_dit` or `tea_cache`):
 
 ```bash
 vllm serve SenseNova/SenseNova-U1-8B-MoT --omni --port 8091 \
-    --cache-backend cache_dit
+    --cache-backend tea_cache
 ```
 
 Or use the convenience script:
@@ -19,8 +19,8 @@ Or use the convenience script:
 cd examples/online_serving/sensenova_u1
 bash run_server.sh
 
-# Cache-DiT
-CACHE_BACKEND=cache_dit bash run_server.sh
+# Cache acceleration (cache_dit or tea_cache)
+CACHE_BACKEND=tea_cache bash run_server.sh
 ```
 
 ### Tensor Parallelism (TP)

examples/online_serving/text_to_speech/README.md

@@ -79,7 +79,7 @@ response = client.audio.speech.create(
 response.stream_to_file("output.wav")
 ```
 
-Streaming PCM output (where supported) — set `stream=true` with `response_format="pcm"`:
+Streaming PCM output (where supported) — set `stream=true`, `stream_format="audio"`, and `response_format="pcm"`:
 
 ```bash
 curl -X POST http://localhost:8091/v1/audio/speech \
@@ -88,6 +88,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "input": "Hello, how are you?",
         "voice": "default",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
@@ -363,6 +364,7 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         \"input\": \"Hello, streaming output from MOSS-TTS-Nano.\",
         \"ref_audio\": \"data:audio/wav;base64,${REF_AUDIO}\",
         \"stream\": true,
+        \"stream_format\": \"audio\",
         \"response_format\": \"pcm\"
     }" --no-buffer | play -t raw -r 48000 -e signed -b 16 -c 1 -
 ```
@@ -528,10 +530,11 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "voice": "vivian",
         "language": "English",
         "stream": true,
+        "stream_format": "audio",
         "response_format": "pcm"
     }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
 ```
-Streaming requires `response_format="pcm"` and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
+Raw PCM streaming requires `stream_format="audio"`, `response_format="pcm"`, and `async_chunk: true` on the stage config (default in `qwen3_tts.yaml`). `speed` is not supported when streaming.
 
 ### Streaming WebSocket
 The `/v1/audio/speech/stream` endpoint accepts text incrementally, splits it at sentence boundaries, and emits one PCM stream per sentence:
@@ -632,7 +635,7 @@ After startup, `/v1/audio/voices` lists `alice`, and `/v1/audio/speech` can use
 ```bash
 python voxcpm2/gradio_demo.py
 ```
-Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Audio is streamed from the OpenAI Speech endpoint with `stream=true`.
+Uses an AudioWorklet-based player adapted from the Qwen3-TTS demo for gap-free playback. Raw PCM audio is streamed from the OpenAI Speech endpoint with `stream=true` and `stream_format="audio"`.
 
 ---
 

examples/online_serving/text_to_speech/cosyvoice3/speech_client.py

@@ -60,6 +60,7 @@ def run_tts(args) -> None:
 
     if args.stream:
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     print(f"Model: {args.model}")

examples/online_serving/text_to_speech/fish_speech/gradio_demo.py

@@ -66,6 +66,8 @@ def build_payload(
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
     }
+    if stream:
+        payload["stream_format"] = "audio"
 
     # Voice cloning: ref_audio takes priority over URL.
     ref_url_stripped = ref_audio_url.strip() if ref_audio_url else ""

examples/online_serving/text_to_speech/fish_speech/speech_client.py

@@ -53,6 +53,7 @@ def run_tts(args) -> None:
 
     if args.stream:
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     print(f"Model: {args.model}")

examples/online_serving/text_to_speech/glm_tts/gradio_demo.py

@@ -64,6 +64,8 @@ def build_payload(
         "stream": stream,
         "max_new_tokens": max_new_tokens,
     }
+    if stream:
+        payload["stream_format"] = "audio"
 
     # Voice cloning: ref_audio takes priority over URL.
     ref_url_stripped = ref_audio_url.strip() if ref_audio_url else ""

examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py

@@ -40,6 +40,9 @@ def run_tts_generation(args) -> None:
         "ref_audio": args.ref_audio,
         "ref_text": args.ref_text,
     }
+    if args.stream:
+        payload["stream_format"] = "audio"
+        payload["response_format"] = "pcm"
     if args.max_new_tokens:
         payload["max_new_tokens"] = args.max_new_tokens
 

examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py

@@ -123,6 +123,7 @@ def run_tts(args) -> None:
         payload["max_new_tokens"] = args.max_new_tokens
     if args.stream:
         payload["stream"] = True
+        payload["stream_format"] = "audio"
         payload["response_format"] = "pcm"
 
     api_url = f"{args.api_base}/v1/audio/speech"

examples/online_serving/text_to_speech/ming_tts/run_curl.sh

@@ -65,6 +65,7 @@ PY
             \"model\": \"${MODEL}\",
             \"input\": \"${TEXT}\",
             \"stream\": true,
+            \"stream_format\": \"audio\",
             \"response_format\": \"pcm\"
         }" "$STREAM_OUTPUT"
         ;;

examples/online_serving/text_to_speech/moss_tts_nano/gradio_demo.py

@@ -335,6 +335,7 @@ def build_payload(
         "ref_audio": encode_audio_to_base64(ref_audio),
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
+        **({"stream_format": "audio"} if stream else {}),
     }
 
 

examples/online_serving/text_to_speech/qwen3_tts/tts_common.py

@@ -93,6 +93,8 @@ def build_payload(
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
     }
+    if stream:
+        payload["stream_format"] = "audio"
     if not stream:
         payload["speed"] = speed
 

examples/online_serving/text_to_speech/voxcpm2/gradio_demo.py

@@ -515,6 +515,8 @@ def create_app(api_base: str):
                 "response_format": "pcm" if stream_enabled else "wav",
                 "stream": stream_enabled,
             }
+            if stream_enabled:
+                payload["stream_format"] = "audio"
 
             # Reference audio for cloning
             ref_url_s = ref_url.strip() if ref_url else ""

pyproject.toml

@@ -41,6 +41,7 @@ dev = [
     "pytest-cov>=4.0.0",
     "pytest-mock>=3.10.0",
     "websockets>=15.0.1",
+    "pytest-subtests>=0.11.0",
     "datasets>=2.14.0",
     "mypy==1.11.1",
     "pre-commit==4.0.1",
@@ -68,7 +69,8 @@ dev = [
     "torchdiffeq>=0.2.5", # for covo_audio Test
     "FlagEmbedding", # for hunyuan_image3 tests
     "kernels<0.15.0", # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
-    "s3tokenizer>=0.3.0" # for tts test
+    "s3tokenizer>=0.3.0", # for tts test
+    "step-audio2>=1.0.0", # Step-Audio2 Token2Wav (flashcosyvoice / hyperpyyaml)
 ]
 
 demo = [
@@ -271,6 +273,7 @@ markers = [
     "skipif_npu: Skip if the num of NPU cards is less than the required",
     "skipif_musa: Skip if the num of MUSA cards is less than the required",
     # more detailed markers
+    "sp: Sequence parallelism tests (multi-GPU)",
     "slow: Slow tests (may skip in quick CI)",
     "benchmark: Benchmark tests",
 ]

recipes/LTX/LTX-2.3.md

@@ -6,21 +6,23 @@
 
 - Vendor: Lightricks
 - Model: `dg845/LTX-2.3-Diffusers`
-- Task: Text-to-video with synchronized audio generation
+- Task: Text-to-video and image-to-video with synchronized audio generation
 - Mode: Online serving (pure diffusion)
 - Maintainer: @oglok
 
 ## When to use this recipe
 
-Use this recipe when you want to serve LTX-2.3 for text-to-video generation
-with audio. The model generates videos up to 20+ seconds at 768x512 resolution
-with 48kHz audio, all from a single text prompt. Requires a GPU with at least
-96GB VRAM due to the 22B parameter transformer (~44GB weights) plus text
+Use this recipe when you want to serve LTX-2.3 for text-to-video or
+image-to-video generation with audio. The model generates videos up to 20+
+seconds at 768x512 resolution with 48kHz audio from text prompts, and can
+condition video generation on an initial image. Start with a 96GB-class GPU for
+validation because the model combines a 22B parameter transformer with text
 encoder, VAE, and vocoder components.
 
 ## References
 
-- Model: <https://huggingface.co/dg845/LTX-2.3-Diffusers>
+- Upstream raw checkpoints: <https://huggingface.co/Lightricks/LTX-2.3>
+- Diffusers-format checkpoint: <https://huggingface.co/dg845/LTX-2.3-Diffusers>
 - Requires `diffusers >= 0.38.0` (install from git: `pip install git+https://github.com/huggingface/diffusers.git`)
 
 ## Serving
@@ -34,7 +36,16 @@ vllm serve dg845/LTX-2.3-Diffusers \
   --stage-init-timeout 600
 ```
 
-### Verification
+For image-to-video:
+
+```bash
+vllm serve dg845/LTX-2.3-Diffusers \
+  --omni \
+  --model-class-name LTX23ImageToVideoPipeline \
+  --stage-init-timeout 600
+```
+
+### T2V Verification
 
 ```bash
 # Health check
@@ -73,19 +84,48 @@ curl -X POST http://localhost:8000/v1/videos \
   -F "guidance_scale=4.0"
 ```
 
+### I2V Verification
+
+Start the server with `--model-class-name LTX23ImageToVideoPipeline`, then pass
+an image reference. The I2V pipeline requires `input_reference` or
+`image_reference` because the first frame is the conditioning input.
+
+```bash
+curl http://localhost:8000/health
+
+curl -X POST http://localhost:8000/v1/videos/sync \
+  -F "prompt=A plush toy astronaut gently waving while the camera slowly pushes in." \
+  -F "input_reference=@/absolute/path/to/reference.png" \
+  -F "model=dg845/LTX-2.3-Diffusers" \
+  -F "num_frames=81" \
+  -F "fps=24" \
+  -F "size=768x512" \
+  -F "num_inference_steps=30" \
+  -F "guidance_scale=4.0" \
+  -F "seed=42" \
+  -o ltx23_i2v.mp4
+```
+
 ### Notes
 
-- Memory usage: Model loads at ~36 GiB, peaks at ~62 GiB during inference
+- Memory usage: run a latest-head serving smoke or DFX benchmark on the target
+  GPU before publishing load/peak VRAM numbers.
+- Checkpoint format: use a Diffusers-format checkpoint such as
+  `dg845/LTX-2.3-Diffusers`; the upstream `Lightricks/LTX-2.3` repository ships
+  raw safetensors and does not contain the subfolder configs required by this
+  pipeline loader.
 - Key flags:
-  - `--stage-init-timeout 600`: Required for the initial `torch.compile` warmup (~90-140 seconds on first request)
-  - `--model-class-name LTX23Pipeline`: Selects the LTX-2.3 pipeline (not LTX-2)
+  - `--stage-init-timeout 600`: Gives the initial `torch.compile` warmup enough time on large checkpoints
+  - `--model-class-name LTX23Pipeline`: Selects the LTX-2.3 text-to-video pipeline (not LTX-2)
+  - `--model-class-name LTX23ImageToVideoPipeline`: Selects the LTX-2.3 image-to-video pipeline
+  - `input_reference=@...`: Uploads the I2V reference image directly
 - Audio: 48kHz AAC via BWE vocoder, automatically synced with video
 - CPU offloading: Text encoder (Gemma-3-12B), connectors, VAE, audio VAE, and vocoder stay on CPU and are moved to GPU only when needed
 - Supported resolutions: 768x512, 512x384 (must be divisible by 32)
 - Frame rate: 24 fps
 - Duration: Controlled by `num_frames` (frames = duration_seconds * 24 + 1)
 - Known limitations:
-  - No image-to-video support yet (LTX23ImageToVideoPipeline is a placeholder)
+  - LTX-2.3 I2V currently supports first-frame image conditioning, matching the LTX image-conditioning contract.
   - Requires `diffusers >= 0.38.0` (not yet on PyPI, install from git)
 
 ## Hardware Support
@@ -99,13 +139,16 @@ curl -X POST http://localhost:8000/v1/videos \
 - OS: Ubuntu 22.04
 - Python: 3.10+
 - Driver / runtime: CUDA 13.0, Driver 580.126.09
-- vLLM version: 0.19.x
-- vLLM-Omni version: 0.19.x
+- vLLM version: Match the PR or release checkout being validated
+- vLLM-Omni version or commit: Use the commit you are deploying from
 
-### Validated configurations
+### Benchmarking
 
-| Duration | Frames | Resolution | Steps | Guidance | Inference Time | Peak VRAM |
-|----------|--------|------------|-------|----------|----------------|-----------|
-| 3s | 81 | 768x512 | 30 | 4.0 | ~110s | ~62 GB |
-| 10s | 241 | 768x512 | 30 | 4.0 | ~130s | ~62 GB |
-| 20s | 481 | 768x512 | 30 | 4.0 | ~420s | ~62 GB |
+Do not copy latency or VRAM numbers between commits or machines. Record them
+only after running a latest-head sweep on the target hardware.
+
+For formal PR benchmarking, reuse
+`tests/dfx/perf/tests/test_ltx2_3_vllm_omni.json` with
+`tests/dfx/perf/scripts/run_diffusion_benchmark.py`. That config captures the
+single-device eager baseline and CFG-parallel=2 case for a small 384x512,
+25-frame, 20-step workload.

recipes/MammothModa2/MammothModa2-Preview.md

@@ -0,0 +1,110 @@
+# MammothModa2-Preview
+
+> MammothModa2-Preview text-to-image generation through the shared offline image example
+
+## Summary
+
+- Vendor: ByteDance Research
+- Model: `bytedance-research/MammothModa2-Preview`
+- Task: Text-to-image generation (AR → DiT two-stage pipeline)
+- Mode: Offline inference
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe to run MammothModa2-Preview text-to-image through the shared
+offline image example (`text_to_image.py`) instead of a model-specific script.
+The generic example formats the AR prompt, drives the AR → DiT stage pipeline,
+and forwards MammothModa2-specific generation parameters through the
+pipeline-declared `extra_body` contract.
+
+MammothModa2's DiT stage consumes its inputs through the multi-stage kwargs
+interface (not `OmniDiffusionRequest`), so its generation knobs
+(`text_guidance_scale`, `cfg_range`, `num_inference_steps`) are passed via
+`--extra-body` rather than the standard `--num-inference-steps` / `--cfg-scale`
+flags. Image size uses the standard `--height` / `--width` flags.
+
+## References
+
+- Upstream model:
+  [`bytedance-research/MammothModa2-Preview`](https://huggingface.co/bytedance-research/MammothModa2-Preview)
+- Related offline example:
+  [`examples/offline_inference/text_to_image/text_to_image.py`](../../examples/offline_inference/text_to_image/text_to_image.py)
+- Declared parameters:
+  [`vllm_omni/model_extras/mammothmodal2_preview.py`](../../vllm_omni/model_extras/mammothmodal2_preview.py)
+- Deploy config:
+  [`vllm_omni/deploy/mammoth_moda2.yaml`](../../vllm_omni/deploy/mammoth_moda2.yaml)
+
+## Hardware Support
+
+The default deploy config runs both the AR and DiT stages on a single GPU
+(`devices: "0"`). The committed `gpu_memory_utilization` split (stage-0 AR `0.5`,
+stage-1 DiT `0.3`) is sized for an ~80 GB GPU. The model also fits on a 48 GB GPU
+after rebalancing the split so the AR weights (~23 GB) leave room for the KV
+cache — see the note under *1x L40S 48GB*.
+
+## GPU
+
+### 1x L40S 48GB
+
+> **48 GB config adjustment:** the committed
+> `vllm_omni/deploy/mammoth_moda2.yaml` uses
+> `gpu_memory_utilization` 0.5 / 0.3 (sized for ~80 GB). To fit on a 48 GB L40S,
+> set the stage-0 (AR) value to `0.8` and the stage-1 (DiT) value to `0.16`
+> before running. (On an ~80 GB GPU, leave the defaults unchanged.)
+
+#### Environment
+
+- OS: Linux
+- Python: Match the repository requirements for your checkout
+- Driver / runtime: NVIDIA CUDA environment with one L40S 48 GB (verified) or an
+  ~80 GB GPU for the default config
+- vLLM version: Match the repository requirements for your checkout
+- vLLM-Omni version or commit: Use the commit you are deploying from
+
+#### Offline Commands
+
+Download the model:
+
+```bash
+hf download bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-Preview
+```
+
+Run text-to-image with the shared offline example from the repository root. The
+deploy config sets `trust_remote_code`, so no extra flag is needed. Forward the
+MammothModa2 generation parameters as a JSON object through `--extra-body`:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model ./MammothModa2-Preview \
+  --stage-configs-path vllm_omni/deploy/mammoth_moda2.yaml \
+  --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
+  --height 1024 \
+  --width 1024 \
+  --extra-body '{"text_guidance_scale": 4.0, "cfg_range": [0.0, 1.0], "num_inference_steps": 50}' \
+  --output mammoth_t2i.png
+```
+
+The `--extra-body` JSON forwards MammothModa2-specific parameters into
+`OmniDiffusionSamplingParams.extra_args`. Keys are filtered against the model's
+declared `extra_body_params` (see
+[`vllm_omni/model_extras/mammothmodal2_preview.py`](../../vllm_omni/model_extras/mammothmodal2_preview.py)),
+so unknown keys for MammothModa2 are silently dropped:
+
+- `text_guidance_scale` — classifier-free guidance scale for the DiT stage
+  (default `9.0`; CFG is active only when `> 1.0`).
+- `cfg_range` — relative step range `[start, end]` over which CFG is applied
+  (default `[0.0, 1.0]`).
+- `num_inference_steps` — number of DiT denoising steps (default `50`).
+
+`--height` and `--width` must be multiples of 16.
+
+#### Verification
+
+The example writes the generated image to the `--output` path. Confirm the file
+exists and is a valid image:
+
+```bash
+ls -lh mammoth_t2i.png
+python -c "from PIL import Image; print(Image.open('mammoth_t2i.png').size)"
+```