Revert "[Perf][Qwen3-TTS] Fuse QKV and gate_up projections in code predictor …"

This reverts commit 6fdc4ca62d.

.buildkite/amd/test-amd-merge.yml

@@ -89,7 +89,7 @@ steps:
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
-        - pytest -s -v tests/diffusion/distributed/test_zimage_parallelism.py
+        - pytest -s -v tests/diffusion/distributed/test_tensor_parallel.py
 
     - label: "Diffusion Model Test"
       agent_pool: mi325_1

.buildkite/amd/test-amd-ready.yml

@@ -35,7 +35,7 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - timeout 20m pytest -s -v tests/e2e/offline_inference/custom_pipeline/ -m "core_model"
+    - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
 
 
 - label: "Voxtral TTS CUDA Unit Test"

.buildkite/cuda/test-merge.yml

@@ -8,14 +8,24 @@ steps:
   - group: ":card_index_dividers: Simple Test"
     depends_on: upload-merge-pipeline
     steps:
-      - label: "Simple · Diffusion & Model Executor Test"
+      - label: "Simple · Model Executor Test"
         commands:
-          - "timeout 20m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        mirror_hardwares: l4_1
+
+      - label: "Simple · Diffusion Test"
+        commands:
+          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        mirror_hardwares: l4_1
+
+      - label: "Simple · Engine&Entrypoints Test"
+        commands:
+          - "timeout 20m pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
         mirror_hardwares: l4_1
 
       - label: "Simple · Other Test"
         commands:
-          - "timeout 20m pytest -sv -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "timeout 20m pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
         mirror_hardwares: l4_1
 
   - group: ":card_index_dividers: Diffusion Test"
@@ -27,12 +37,6 @@ steps:
           - pytest -sv tests/diffusion/lora/ -m 'advanced_model and cuda' --run-level "advanced_model"
         mirror_hardwares: l4_1
 
-      - label: "Diffusion · Offloader Test"
-        timeout_in_minutes: 40
-        commands:
-          - timeout 40m pytest -s -v tests/diffusion/offloader -m "advanced_model and cuda"
-        mirror_hardwares: l4_1
-
   - label: "Distributed Test"
     depends_on: upload-merge-pipeline
     timeout_in_minutes: 60
@@ -135,39 +139,32 @@ steps:
         timeout_in_minutes: 50
         commands:
           - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.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
+        mirror_hardwares: h100_2
+
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        source_file_dependencies:
+          - tests/assets/minicpmo_4_5/response_required_16k.wav
+          - tests/assets/minicpmo_4_5/soft_interrupt_16k.wav
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
+          - tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+          - tests/helpers/minicpmo_4_5_duplex.py
+          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
+          - vllm_omni/experimental/fullduplex/
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
+          - vllm_omni/core/sched/
+          - vllm_omni/outputs/multimodal_accumulation.py
+          - vllm_omni/engine/stage_init_utils.py
+          - vllm_omni/engine/async_omni_engine.py
+          - vllm_omni/entrypoints/cli/serve.py
+        timeout_in_minutes: 50
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+        mirror_hardwares: h100_2
 
       - label: "TTS · VoxCPM2 Test"
         source_file_dependencies:
@@ -185,22 +182,6 @@ steps:
             "
         mirror_hardwares: l4_1
 
-      - label: "TTS · SoulX-Singer Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_soulxsinger.py
-          - vllm_omni/diffusion/models/soulx_singer/
-          - vllm_omni/deploy/soulxsinger_svc.yaml
-          - vllm_omni/deploy/soulxsinger_svs.yaml
-        timeout_in_minutes: 30
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/offline_inference/test_soulxsinger.py -k 'svc or precomputed' -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
-
       - label: "Diffusion · Qwen Image Edit Test"
         source_file_dependencies:
           - tests/e2e/online_serving/test_qwen_image_edit.py

.buildkite/cuda/test-nightly.yml

@@ -15,7 +15,13 @@ steps:
       - label: ":full_moon: Omni · Function Test with H100"
         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
+          - 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 --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 Duplex Test"
+        timeout_in_minutes: 50
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni" --run-level "full_model"
         mirror_hardwares: h100_2
 
       - label: ":full_moon: Omni · Doc Test with L4"
@@ -421,6 +427,12 @@ steps:
           - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2v_matches_diffusers' -m full_model --run-level full_model
         mirror_hardwares: h100_2
 
+      - label: ":full_moon: Diffusion X2V · MiniMax H3 T2VA Accuracy Test"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/e2e/accuracy/minimax_h3/test_minimax_h3_t2va_similarity.py -m full_model --run-level full_model
+        mirror_hardwares: h100_4
+
       - label: ":full_moon: Diffusion X2V · Other Accuracy Test"
         timeout_in_minutes: 180
         commands:
@@ -449,36 +461,47 @@ steps:
         mirror_hardwares: h100_2
 
   - group: ":card_index_dividers: Diffusion Test"
+    key: nightly-diffusion-test-group
     depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "diffusion-x2iat-test"
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
     steps:
+      - label: ":full_moon: Diffusion · Distributed Test with L4"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4' --run-level "full_model"
+        mirror_hardwares: l4_4
+
+      - label: ":full_moon: Diffusion · Distributed Test with H100"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100' --run-level "full_model"
+        mirror_hardwares: h100_4
 
       - label: ":full_moon: Diffusion · Offloader Test"
-        timeout_in_minutes: 40
+        timeout_in_minutes: 60
         commands:
-          - timeout 40m pytest -s -v tests/diffusion/offloader -m "full_model and cuda"
+          - pytest -sv tests/diffusion/offloader -m "full_model and cuda"
         mirror_hardwares: l4_1
 
-      - label: ":full_moon: Diffusion · Quantization Test with H100"
+      - label: ":full_moon: Diffusion · Quantization Test with H100 · Single-GPU"
+        timeout_in_minutes: 60
         commands:
-          - timeout 60m pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100' --run-level "full_model"
+          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100' --run-level "full_model"
         mirror_hardwares: h100_1
 
+      - label: ":full_moon: Diffusion · MiniMax H3 FP8 Quality Test with H100 · 2-GPU"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py -m 'full_model and cuda and H100' --run-level "full_model"
+        mirror_hardwares: h100_2
+
       - label: ":full_moon: Diffusion · Quantization Test with L4"
         timeout_in_minutes: 60
         commands:
+          - pip install "vllm-gguf-plugin>=0.0.3"
           - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and L4' --run-level "full_model"
         mirror_hardwares: l4_1
 
-      - label: ":full_moon: Diffusion · Wan VAE Spatial Shard Correctness Test"
-        timeout_in_minutes: 60
-        commands:
-          - >-
-            pytest -sv tests/diffusion/distributed/test_wan_spatial_shard.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
-            --run-level "full_model"
-        mirror_hardwares: h100_2
-
   - label: ":bar_chart: Testcase Statistics"
     key: nightly-testcase-statistics
     timeout_in_minutes: 120

.buildkite/cuda/test-ready.yml

@@ -8,22 +8,26 @@ steps:
   - group: ":card_index_dividers: Simple Test"
     depends_on: upload-ready-pipeline
     steps:
-      - label: "Simple · Diffusion & Model Executor Test"
+      - label: "Simple · Model Executor Test"
         commands:
-          - "timeout 20m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        mirror_hardwares: l4_1
+
+      - label: "Simple · Diffusion Test"
+        commands:
+          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        mirror_hardwares: l4_1
+
+      - label: "Simple · Engine&Entrypoints Test"
+        commands:
+          - "timeout 20m pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
         mirror_hardwares: l4_1
 
       - label: "Simple · Other Test"
         commands:
-          - "timeout 20m pytest -sv -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "timeout 20m pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
         mirror_hardwares: l4_1
 
-  - label: "Custom Pipeline Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - timeout 20m pytest -s -v tests/e2e/offline_inference/custom_pipeline/ -m "core_model"
-    mirror_hardwares: l4_1
-
   - group: ":card_index_dividers: Diffusion Test"
     depends_on: upload-ready-pipeline
     steps:
@@ -104,11 +108,34 @@ steps:
   - group: ":card_index_dividers: E2E Test"
     depends_on: upload-ready-pipeline
     steps:
+      - label: "Custom Pipeline Test"
+        source_file_dependencies:
+          - tests/e2e/features/custom_pipeline/
+          - tests/e2e/features/helpers/custom_pipeline.py
+          - vllm_omni/diffusion/
+          - vllm_omni/engine/
+          - vllm_omni/entrypoints/
+          - vllm_omni/inputs/
+          - vllm_omni/outputs/
+        commands:
+          - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "RLHF · VeRL-Omni E2E Test"
         source_file_dependencies:
-          - tests/e2e/offline_inference/rlhf_test/test_verl_omni_e2e.py
-          - tests/e2e/offline_inference/custom_pipeline/
+          - tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
+          - tests/e2e/features/helpers/verl_omni_server.py
+          - tests/e2e/features/helpers/custom_pipeline.py
           - vllm_omni/diffusion/
           - vllm_omni/engine/
           - vllm_omni/entrypoints/
@@ -117,7 +144,7 @@ steps:
         commands:
           - |
             timeout 20m bash -c "
-              pytest -s -v tests/e2e/offline_inference/rlhf_test/test_verl_omni_e2e.py
+              pytest -s -v tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
             "
         mirror_hardwares: l4_1
 
@@ -143,34 +170,7 @@ steps:
             timeout 20m bash -c "
               pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py -m 'core_model' --run-level 'core_model'
             "
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 2
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
+        mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Online Test"
         source_file_dependencies:
@@ -181,34 +181,34 @@ steps:
             timeout 20m bash -c "
               pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5.py -m 'core_model' --run-level 'core_model'
             "
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 2
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
+        mirror_hardwares: h100_2
+
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        source_file_dependencies:
+          - tests/assets/minicpmo_4_5/response_required_16k.wav
+          - tests/assets/minicpmo_4_5/soft_interrupt_16k.wav
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
+          - tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+          - tests/helpers/minicpmo_4_5_duplex.py
+          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
+          - vllm_omni/experimental/fullduplex/
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
+          - vllm_omni/core/sched/
+          - vllm_omni/outputs/multimodal_accumulation.py
+          - vllm_omni/engine/stage_init_utils.py
+          - vllm_omni/engine/async_omni_engine.py
+          - vllm_omni/entrypoints/cli/serve.py
+        commands:
+          - |
+            timeout 20m bash -c "
+              pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'core_model and cuda' --run-level 'core_model'
+            "
+        mirror_hardwares: h100_2
 
       - label: "TTS · VoxCPM2 Test"
         source_file_dependencies:
@@ -344,13 +344,3 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_model'
             "
         mirror_hardwares: h100_1
-
-      - label: "Diffusion · GGUF Plugin Test"
-        soft_fail: true
-        source_file_dependencies:
-          - tests/plugin/test_gguf_diffusion.py
-          - vllm_omni/diffusion/model_loader
-        commands:
-          - pip install "vllm-gguf-plugin>=0.0.3"
-          - timeout 30m pytest -sv tests/plugin/test_gguf_diffusion.py
-        mirror_hardwares: l4_1

.buildkite/intel/pipeline-intel.yml

@@ -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: "v0.25.0"
+          VLLM_VERSION: "v0.26.0"
         priority: 100
         timeout_in_minutes: 60
         soft_fail: false

.buildkite/npu/test-npu-nightly.yml

@@ -95,6 +95,80 @@ steps:
             buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             exit $$EXIT
+  - group: ":card_index_dividers: Diffusion I2I Model Test"
+    key: nightly-HunyuanImage3-x2i-test-group
+    depends_on: upload-nightly-pipeline
+    if: >-
+      build.env("NIGHTLY") == "1" ||
+      build.pull_request.labels includes "nightly-test" ||
+      build.pull_request.labels includes "diffusion-x2iat-test"
+    steps:
+      - label: ":full_moon: HunyuanImage3-DIT · Accuracy Test"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_4
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          HUNYUAN_IMAGE3_MODEL: "tencent/HunyuanImage-3.0-Instruct"
+          HUNYUAN_IMAGE3_DEVICES: "0,1,2,3"
+          DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
+        commands:
+          - pytest -s -v tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py -m "full_model and A3" --run-level full_model
+  - group: ":card_index_dividers: Diffusion Benchmark Perf Test"
+    key: nightly-diffusion-benchmark-group
+    depends_on: upload-nightly-pipeline
+    if: >-
+      build.env("NIGHTLY") == "1" ||
+      build.pull_request.labels includes "nightly-test" ||
+      build.pull_request.labels includes "diffusion-x2iat-test"
+    steps:
+      - label: ":full_moon: Diffusion Perf · HunyuanImage3 TP=2 CFGP=2"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_4
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          HF_TOKEN: "${HF_TOKEN}"
+          DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m A3 --test-config-file tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+      - label: ":full_moon: Diffusion Perf · HunyuanImage3 TP=2 SP=2"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_4
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          HF_TOKEN: "${HF_TOKEN}"
+          DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m A3 --test-config-file tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+      - label: ":full_moon: Diffusion Perf · HunyuanImage3 TP=4"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_4
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          HF_TOKEN: "${HF_TOKEN}"
+          DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m A3 --test-config-file tests/dfx/perf/tests/test_hunyuan_image_tp4.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
   - group: ":card_index_dividers: Diffusion X2V Model Test"
     key: nightly-diffusion-x2v-group
     depends_on: upload-nightly-pipeline
@@ -103,6 +177,14 @@ steps:
       build.pull_request.labels includes "nightly-test" ||
       build.pull_request.labels includes "diffusion-x2v-test"
     steps:
+      - label: ":full_moon: Diffusion X2V · Function Test"
+        mirror_hardwares: a2b3_npu_4
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k npu_i2v --run-level "full_model"
       - label: ":full_moon: Diffusion X2V · Perf Test"
         key: nightly-diffusion-x2v-performance
         timeout_in_minutes: 180

README.md

@@ -16,14 +16,10 @@ Easy, fast, and cheap omni-modality model serving for everyone
 ---
 
 *Latest News* 🔥
+- [2026/08] We released [0.26.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.26.0) - aligned with the vLLM 0.26 release line, featuring [MiniMax H3](recipes/MiniMaxAI/MiniMax-H3.md) joint video/audio generation, an experimental full-duplex realtime runtime for [MiniCPM-o 4.5](recipes/OpenBMB/MiniCPM-o-4_5.md), distributed layerwise diffusion offload, and broader model, hardware, streaming, TTS, and quantization support.
 - [2026/07] We released [0.24.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.24.0) - aligned with the vLLM 0.24 release line, expanding production-ready coverage across TTS, speech, diffusion, image/video generation, and robot-policy serving, with major Omni stage runtime refactoring, diffusion request-level batching, async output materialization, quantization/cache/memory improvements, and broad CUDA/ROCm/XPU/NPU support.
-- [2026/06] We released [0.22.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.22.0) - an **omnimodal world-model** release aligned with vLLM 0.22, featuring [Nvidia Cosmos3](recipes/cosmos3/Cosmos3-Nano.md)/DreamZero world model support, expanded quantization coverage across Blackwell/NPU/XPU, TTS production improvements, new models including MiniCPM-o 4.5, MOSS-TTS, and Lance, plus RL integration with [VeRL-Omni](https://github.com/verl-project/verl-omni).
-- [2026/05] We released [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0) - refreshes the serving/runtime stack for large-scale omni workloads, and improves diffusion model performance, quantization, and hardware readiness across CUDA, ROCm, MUSA, NPU, and XPU backends.
-- [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments.
+- [2026/06] Starting with [0.14.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.14.0), vLLM-Omni publishes a stable release aligned with every even-numbered upstream vLLM minor version. [0.16.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.16.0), [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0), [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0), and [0.22.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.22.0) continued this cadence, expanding omni and world-model support with [NVIDIA Cosmos3](recipes/cosmos3/Cosmos3-Nano.md) and DreamZero, adding models such as MiniCPM-o 4.5, MOSS-TTS, and Lance, and advancing TTS, diffusion, distributed execution, quantization, RL integration through [VeRL-Omni](https://github.com/verl-project/verl-omni), and CUDA/ROCm/MUSA/NPU/XPU coverage.
 - [2026/03] Check out our first public [project deepdive](https://youtu.be/sgwNfsNnR9I) at the vLLM Hong Kong Meetup!
-- [2026/03] **[vllm-omni-skills](https://github.com/hsliuustc0106/vllm-omni-skills)** is a community-driven collection of AI assistant skills that help developers work with vLLM-Omni more effectively. These skills can be used with popular agentic AI coding assistants like **Cursor IDE**, **Claude**, **Codex**, and more.
-- [2026/02] We released [0.16.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.16.0) - A major alignment + capability release that rebases onto **upstream vLLM v0.16.0** and significantly expands performance, distributed execution, and production readiness across **Qwen3-Omni / Qwen3-TTS**, **Bagel**, **MiMo-Audio**, **GLM-Image** and the **Diffusion (DiT) image/video stack**—while also improving platform coverage (CUDA / ROCm / NPU / XPU), CI quality, and documentation.
-- [2026/02] We released [0.14.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.14.0) - This is the first **stable release** of vLLM-Omni that expands Omni’s diffusion / image-video generation and audio / TTS stack, improves distributed execution and memory efficiency, and broadens platform/backend coverage (GPU/ROCm/NPU/XPU). It also brings meaningful upgrades to serving APIs, profiling & benchmarking, and overall stability. Please check our latest [paper](https://arxiv.org/abs/2602.02204) for architecture design and performance results.
 - [2025/11] vLLM community officially released [vllm-project/vllm-omni](https://github.com/vllm-project/vllm-omni) in order to support omni-modality models serving.
 
 ---
@@ -55,12 +51,13 @@ vLLM-Omni is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
+- Full-duplex realtime serving with streaming audio input and output (experimental)
 
 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:
 
-- **Omni-modality models** (e.g. Qwen3-Omni, Cosmos3, HunyuanImage, BAGEL)
+- **Omni-modality models** (e.g. Qwen3-Omni, MiniCPM-o 4.5, Cosmos3, HunyuanImage, BAGEL)
 - **TTS models** (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
-- **Diffusion models** — image, video, and audio generation (e.g. Qwen-Image, Wan2.2, FLUX)
+- **Diffusion models** — image, video, and audio generation (e.g. MiniMax H3, Qwen-Image, Wan2.2, FLUX)
 - **Robot-policy and action models** (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)
 
 ## Getting Started

apps/ComfyUI-vLLM-Omni/README.md

@@ -47,7 +47,7 @@ If no, check your shell running the ComfyUI process. There may be some error mes
 This extension offers the following nodes based on the output modalities (at **ComfyUI sidebar -> Node Library**):
 
 - **Generate Image** for text-to-image and image-to-image tasks
-- **Generate Video** for text-to-video and image-to-video tasks
+- **Generate Video** for text-to-video, first-frame/image-to-video, and reference-conditioned video
 - **Multimodality Understanding** for multimodality-to-text and multimodality-to-audio tasks
 - **TTS** and **TTS Voice Clone** for TTS tasks
 
@@ -105,7 +105,7 @@ You can configure per-stage sampling parameters for multi-stage models.
 > [!TIP]
 > The node automatically choose text-to-image or image-to-image API endpoints depending on whether you connect an image input or not.
 
-### Text-to-video and image-to-video generation (e.g., Wan)
+### Text-to-video and image-to-video generation (e.g., Wan, MiniMax-H3)
 
 (Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni Video Generation**)
 
@@ -117,7 +117,11 @@ You can configure per-stage sampling parameters for multi-stage models.
 </p>
 
 > [!TIP]
-> The node automatically choose text-to-video or image-to-video API endpoints depending on whether you connect an image input or not.
+> Connect a **frame** image for first-frame / image-to-video (e.g., Wan I2V, MiniMax-H3 FL2VA).
+>
+> For reference-conditioned generation (MiniMax-H3 Ref2VA), connect a **Video References** node instead.
+>
+> Do not use `frame` and `references` together. Task routing is automatic from which inputs you connect.
 
 ### TTS (e.g., Qwen TTS series)
 

apps/ComfyUI-vLLM-Omni/__init__.py

@@ -15,11 +15,13 @@ from .comfyui_vllm_omni.nodes import (
     VLLMOmniDiffusionSampling,
     VLLMOmniGenerateImage,
     VLLMOmniGenerateVideo,
+    VLLMOmniMiniMaxH3Params,
     VLLMOmniQwenTTSParams,
     VLLMOmniRemoteLoRA,
     VLLMOmniSamplingParamsList,
     VLLMOmniTTS,
     VLLMOmniUnderstanding,
+    VLLMOmniVideoReferences,
     VLLMOmniVoiceClone,
     VLLMOmniWanParams,
 )
@@ -32,6 +34,7 @@ NODE_CLASS_MAPPINGS = {
     "VLLMOmniUnderstanding": VLLMOmniUnderstanding,
     "VLLMOmniTTS": VLLMOmniTTS,
     "VLLMOmniVoiceClone": VLLMOmniVoiceClone,
+    "VLLMOmniVideoReferences": VLLMOmniVideoReferences,
     # === Params ===
     "VLLMOmniARSampling": VLLMOmniARSampling,
     "VLLMOmniDiffusionSampling": VLLMOmniDiffusionSampling,
@@ -39,6 +42,7 @@ NODE_CLASS_MAPPINGS = {
     "VLLMOmniRemoteLoRA": VLLMOmniRemoteLoRA,
     "VLLMOmniQwenTTSParams": VLLMOmniQwenTTSParams,
     "VLLMOmniWanParams": VLLMOmniWanParams,
+    "VLLMOmniMiniMaxH3Params": VLLMOmniMiniMaxH3Params,
 }
 
 # A dictionary that contains the friendly/humanly readable titles for the nodes
@@ -49,6 +53,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "VLLMOmniUnderstanding": "Multimodality Understanding",
     "VLLMOmniTTS": "TTS (Text to Speech)",
     "VLLMOmniVoiceClone": "TTS Voice Cloning",
+    "VLLMOmniVideoReferences": "Video References",
     # === Params ===
     "VLLMOmniARSampling": "AR Sampling Params",
     "VLLMOmniDiffusionSampling": "Diffusion Sampling Params",
@@ -56,6 +61,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "VLLMOmniRemoteLoRA": "LoRA",
     "VLLMOmniQwenTTSParams": "Qwen TTS Params",
     "VLLMOmniWanParams": "Wan Video Params",
+    "VLLMOmniMiniMaxH3Params": "MiniMax-H3 Video Params",
 }
 
 WEB_DIRECTORY = "./web"

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

@@ -10,7 +10,9 @@ from .utils.types import (
     AudioFormat,
     AutoregressionSamplingParams,
     DiffusionSamplingParams,
+    MiniMaxH3ModelSpecificParams,
     QwenTTSModelSpecificParams,
+    VideoReferences,
     WanModelSpecificParams,
 )
 from .utils.validators import (
@@ -163,7 +165,8 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
                 "num_frames": ("INT", {"default": 41, "min": 1}),
             },
             "optional": {
-                "image": ("IMAGE",),
+                "frame": ("IMAGE",),
+                "references": ("VIDEO_REFERENCES",),
                 "sampling_params": ("SAMPLING_PARAMS",),
                 "lora": ("REMOTE_LORA",),
                 "model_params": ("VIDEO_PARAMS",),
@@ -174,6 +177,15 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
     RETURN_NAMES = ("video",)
     FUNCTION = "generate"
 
+    @classmethod
+    def VALIDATE_INPUTS(cls, url, model, frame=None, references=None, **_kwargs) -> str | Literal[True]:
+        base = super().VALIDATE_INPUTS(url, model)
+        if base is not True:
+            return base
+        if frame is not None and references is not None:
+            return "Provide only one of frame or references, not both."
+        return True
+
     async def generate(
         self,
         url: str,
@@ -184,7 +196,8 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
         fps: int,
         num_frames: int,
         negative_prompt: str | None = None,
-        image: torch.Tensor | None = None,
+        frame: torch.Tensor | None = None,
+        references: dict | None = None,
         sampling_params: dict | list[dict] | None = None,
         model_params: dict | None = None,
         lora: dict | None = None,
@@ -207,14 +220,16 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
 
         if sampling_params is not None:
             sampling_params.pop("type", None)  # internal fields
-        if model_params is not None:
-            model_params.pop("type", None)  # internal fields
+        # model_params["type"] is also an internal field, but is used in api_client
+        # if model_params is not None:
+        #     model_params.pop("type", None)  # internal fields
 
         client = VLLMOmniClient(url)
         output = await client.generate_video(
             model=model,
             prompt=prompt,
-            image=image,  # image present => i2v, absent => t2v
+            frame=frame,  # frame present => fl2va / Wan I2V
+            references=references,
             width=width,
             height=height,
             num_frames=num_frames,
@@ -734,3 +749,78 @@ class VLLMOmniWanParams:
 
     def get_params(self, **kwargs):
         return (WanModelSpecificParams(kwargs),)
+
+
+class VLLMOmniMiniMaxH3Params:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "audio_flow_shift": (
+                    "FLOAT",
+                    {"default": 3.0, "min": 0.0, "max": 100.0, "step": 0.1},
+                ),
+                "flow_shift": (
+                    "FLOAT",
+                    {"default": 12.0, "min": 0.0, "max": 100.0, "step": 0.1},
+                ),
+            }
+        }
+
+    RETURN_TYPES = ("VIDEO_PARAMS",)
+    RETURN_NAMES = ("MiniMax-H3 video params",)
+    FUNCTION = "get_params"
+    CATEGORY = "vLLM-Omni/Video Params"
+
+    def get_params(self, **kwargs):
+        params = MiniMaxH3ModelSpecificParams(kwargs)
+        params["type"] = "minimax_h3"
+        return (params,)
+
+
+class VLLMOmniVideoReferences:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {},
+            "optional": {
+                "image_1": ("IMAGE",),
+                "image_2": ("IMAGE",),
+                "audio_1": ("AUDIO",),
+                "audio_2": ("AUDIO",),
+                "video_1": ("VIDEO",),
+                "video_2": ("VIDEO",),
+            },
+        }
+
+    RETURN_TYPES = ("VIDEO_REFERENCES",)
+    RETURN_NAMES = ("references",)
+    FUNCTION = "get_references"
+    CATEGORY = "vLLM-Omni"
+
+    def get_references(
+        self,
+        image_1: torch.Tensor | None = None,
+        image_2: torch.Tensor | None = None,
+        audio_1: AudioInput | None = None,
+        audio_2: AudioInput | None = None,
+        video_1: VideoInput | None = None,
+        video_2: VideoInput | None = None,
+        **kwargs,
+    ):
+        if kwargs:
+            logger.info("Uncaught kwargs: %s", kwargs)
+        refs = VideoReferences()
+        if image_1 is not None:
+            refs["image_1"] = image_1
+        if image_2 is not None:
+            refs["image_2"] = image_2
+        if audio_1 is not None:
+            refs["audio_1"] = audio_1
+        if audio_2 is not None:
+            refs["audio_2"] = audio_2
+        if video_1 is not None:
+            refs["video_1"] = video_1
+        if video_2 is not None:
+            refs["video_2"] = video_2
+        return (refs,)

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

@@ -24,6 +24,7 @@ from .format import (
     image_tensor_to_base64,
     image_tensor_to_png_bytes,
     video_to_base64,
+    video_to_bytes,
 )
 from .logger import get_logger, pretty_printer
 from .models import lookup_model_spec
@@ -63,7 +64,11 @@ async def url_bytes(session: aiohttp.ClientSession, url: str, verb: str = "get",
 
 class VLLMOmniClient:
     def __init__(
-        self, base_url: str, timeout: float | None = None, poll_interval: float = 5.0, max_poll_duration: float = 60 * 5
+        self,
+        base_url: str,
+        timeout: float | None = None,
+        poll_interval: float = 5.0,
+        max_poll_duration: float = 60 * 30,
     ):
         self.base_url = base_url
         self.timeout = aiohttp.ClientTimeout(total=timeout)
@@ -261,12 +266,17 @@ class VLLMOmniClient:
         num_frames: int,
         fps: int,
         negative_prompt: str | None = None,
-        image: torch.Tensor | None = None,
+        frame: torch.Tensor | None = None,
+        references: dict | None = None,
         sampling_params: dict | None = None,
         model_params: dict | None = None,
         lora: dict | None = None,
-        **extra_body,
+        **extra_params,
     ) -> VideoInput:
+        if frame is not None and references is not None:
+            raise ValueError("Provide only one of frame or references, not both.")
+
+        # === regular payload fields ===
         form = aiohttp.FormData()
         form.add_field("model", model)
         form.add_field("prompt", prompt)
@@ -279,23 +289,91 @@ class VLLMOmniClient:
         if sampling_params is not None:
             for k, v in sampling_params.items():
                 form.add_field(k, str(v))
-        if model_params is not None:
-            for k, v in model_params.items():
-                form.add_field(k, str(v))
         if lora is not None:
             form.add_field("lora", json.dumps(lora, ensure_ascii=False))
-        if extra_body:
-            form.add_field("extra_body", json.dumps(extra_body, ensure_ascii=False))
 
-        if image is not None:
+        # === multimodal inputs (first-last-frames, references, etc.) ===
+        input_reference_image: torch.Tensor | None = None
+        audio_reference: AudioInput | None = None
+        reference_videos: list[VideoInput] = []
+        video_task: str | None = None
+
+        if frame is not None:
+            input_reference_image = frame
+            video_task = "fl2va"
+        elif references is not None:
+            images = [references[k] for k in ("image_1", "image_2") if k in references and references[k] is not None]
+            audios = [references[k] for k in ("audio_1", "audio_2") if k in references and references[k] is not None]
+            videos = [references[k] for k in ("video_1", "video_2") if k in references and references[k] is not None]
+
+            if not images and not audios and not videos:
+                raise ValueError("references is empty; connect at least one image/audio/video.")
+
+            if videos:
+                if images or audios:
+                    raise ValueError(
+                        "Video references cannot be combined with image or audio references. "
+                        "Connect only video_1/video_2 for multi-video Ref2VA."
+                    )
+                reference_videos = videos
+                video_task = "ref2va"
+            elif len(images) == 1 and len(audios) == 1:
+                input_reference_image = images[0]
+                audio_reference = audios[0]
+                video_task = "ref2va"
+            else:
+                raise ValueError(
+                    "Invalid references combination. Supported modes: "
+                    "(1) one or more videos only, or (2) exactly one image and one audio."
+                )
+        else:
+            video_task = "t2va"
+
+        if input_reference_image is not None:
             image_filename = "image.png"  # Required for multipart form
             form.add_field(
                 "input_reference",
-                image_tensor_to_png_bytes(image, image_filename),
+                image_tensor_to_png_bytes(input_reference_image, image_filename),
                 filename=image_filename,
                 content_type="image/png",
             )
 
+        if audio_reference is not None:
+            form.add_field(
+                "audio_reference",
+                json.dumps({"audio_url": audio_to_base64(audio_reference)}, ensure_ascii=False),
+            )
+
+        for idx, video in enumerate(reference_videos, start=1):
+            video_filename = f"reference_{idx}.mp4"
+            form.add_field(
+                "input_references",
+                video_to_bytes(video, video_filename),
+                filename=video_filename,
+                content_type="video/mp4",
+            )
+
+        # === model specific params. Either use a specialized builder, or add flattened fields as-is ===
+        if model_params is not None:
+            model_params = dict(model_params)
+            model_params.pop("type", None)
+
+        spec, _ = lookup_model_spec(model)
+        params_builder = spec.get("params_builder") if spec else None
+        if params_builder is not None:
+            form_fields = params_builder(
+                model_params or {},
+                extra_params={**extra_params, "task": video_task},
+            )
+            for k, v in form_fields.items():
+                form.add_field(k, v if isinstance(v, str) else str(v))
+        else:
+            if model_params is not None:
+                for k, v in model_params.items():
+                    form.add_field(k, str(v))
+            if extra_params:
+                form.add_field("extra_params", json.dumps(extra_params, ensure_ascii=False))
+
         async with aiohttp.ClientSession(timeout=self.timeout) as session:
             # Start the video generation job
             url = f"{self.base_url}/videos"

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

@@ -1,7 +1,12 @@
+import json
 import re
+from typing import Any
 
+from .logger import get_logger
 from .types import Modality, ModelMode, Spec
 
+logger = get_logger(__name__)
+
 
 def _bagel_payload_preprocessor(payload: dict) -> dict:
     try:
@@ -29,6 +34,30 @@ def _qwen25_payload_preprocessor(payload: dict) -> dict:
     return payload
 
 
+def _minimaxh3_params_builder(
+    model_params: dict[str, Any],
+    *,
+    extra_params: dict[str, Any],
+) -> dict[str, Any]:
+    """Build multipart form fields for MiniMax-H3 from model_params + routed task."""
+    params = dict(model_params)
+    params.pop("type", None)
+    form_fields: dict[str, Any] = {}
+    merged_extra_params: dict[str, Any] = dict(extra_params or {})
+
+    if "flow_shift" in params:
+        form_fields["flow_shift"] = params.pop("flow_shift")
+    for key in ("audio_flow_shift",):
+        if key in params:
+            merged_extra_params[key] = params.pop(key)
+    if params:
+        logger.warning("Unused MiniMax-H3 model params ignored: %s", sorted(params))
+
+    if merged_extra_params:
+        form_fields["extra_params"] = json.dumps(merged_extra_params, ensure_ascii=False)
+    return form_fields
+
+
 _MODEL_PIPELINE_SPECS: dict[str, Spec] = {
     r"BAGEL-7B-MoT": {
         "stages": [
@@ -71,6 +100,21 @@ _MODEL_PIPELINE_SPECS: dict[str, Spec] = {
             }
         ],
     },
+    r"MiniMax-H3(/FL2VA|/Ref2VA)?": {
+        "stages": ["diffusion"],
+        "modes": [
+            {
+                "mode": ModelMode.VIDEO_GENERATION,
+                "input_modalities": [
+                    Modality.TEXT,
+                    Modality.IMAGE,
+                    Modality.VIDEO,
+                    Modality.AUDIO,
+                ],
+            }
+        ],
+        "params_builder": _minimaxh3_params_builder,
+    },
 }
 # Convert dict keys to regex patterns
 MODEL_PIPELINE_SPECS: dict[re.Pattern, Spec] = {}
@@ -80,12 +124,10 @@ del _MODEL_PIPELINE_SPECS
 
 
 def lookup_model_spec(model: str) -> tuple[Spec | None, str | None]:
-    try:
-        last_component = model.rstrip("/").rsplit("/", 1)[-1]
-    except IndexError:
-        last_component = model
+    normalized = model.rstrip("/").lstrip("/")
+    lookup_key = normalized.split("/", 1)[-1]
     for pattern, spec in MODEL_PIPELINE_SPECS.items():
-        if pattern.search(last_component):
+        if pattern.search(lookup_key):
             return spec, pattern.pattern
     return None, None
 
@@ -101,6 +143,9 @@ if __name__ == "__main__":
         "Qwen/Qwen3-Omni-30B-A3B-Instruct",
         "Custom/Path/UnknownModel-Instruct",
         "Not/Matching/Anything",
+        "/path/to/MiniMax-H3/FL2VA",
+        "/path/to/MiniMax-H3/Ref2VA",
+        "MiniMaxAI/MiniMax-H3",
     ]
 
     test_payload = {"messages": [{"role": "user", "content": "prompt"}]}
@@ -112,7 +157,13 @@ if __name__ == "__main__":
             if preprocessor := spec.get("payload_preprocessor"):
                 result = preprocessor(test_payload)
                 print(f"✓ {path:<40} → {result}")
+            elif params_builder := spec.get("params_builder"):
+                result = params_builder(
+                    {"audio_flow_shift": 3.0, "flow_shift": 12.0, "type": "minimax_h3"},
+                    extra_params={"task": "t2va"},
+                )
+                print(f"✓ {path:<40} → {result}")
             else:
-                print(f"✓ {path:<40} → No preprocessor")
+                print(f"✓ {path:<40} → No preprocessor/params_builder")
         else:
             print(f"✗ {path:<40} → No match")

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

@@ -1,12 +1,8 @@
 from collections.abc import Callable
 from enum import Enum, auto
-from typing import (
-    Any,
-    Literal,
-    NotRequired,
-    TypeAlias,
-    TypedDict,
-)
+from typing import Any, Literal, TypeAlias
+
+from typing_extensions import NotRequired, TypedDict
 
 AudioFormat: TypeAlias = Literal["mp3", "opus", "aac", "flac", "wav", "pcm"]
 
@@ -27,6 +23,14 @@ class WanModelSpecificParams(dict):
     pass
 
 
+class MiniMaxH3ModelSpecificParams(dict):
+    pass
+
+
+class VideoReferences(dict):
+    pass
+
+
 class ModelMode(Enum):
     IMAGE_GENERATION = auto()
     VIDEO_GENERATION = auto()
@@ -47,9 +51,11 @@ class ModelModeSpec(TypedDict):
 
 
 PayloadPreprocessor: TypeAlias = Callable[[dict[str, Any]], dict[str, Any]]
+ParamsBuilder: TypeAlias = Callable[..., dict[str, Any]]
 
 
 class Spec(TypedDict):
     stages: list[Literal["diffusion", "autoregression"]]
     modes: list[ModelModeSpec]
     payload_preprocessor: NotRequired[PayloadPreprocessor]
+    params_builder: NotRequired[ParamsBuilder]

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

@@ -1,513 +1 @@
-{
-    "id": "3aea1d08-7de5-4997-ba71-b2f66ad72a21",
-    "revision": 0,
-    "last_node_id": 14,
-    "last_link_id": 28,
-    "nodes": [
-        {
-            "id": 10,
-            "type": "MarkdownNote",
-            "pos": [
-                227.99306819575278,
-                -231.24143306069843
-            ],
-            "size": [
-                240.3326478922474,
-                136.3505820791881
-            ],
-            "flags": {},
-            "order": 0,
-            "mode": 0,
-            "inputs": [],
-            "outputs": [],
-            "title": "Note: Task and Input",
-            "properties": {},
-            "widgets_values": [
-                "vLLM-Omni nodes are categorized based on the output modality. The \"Generate Video\" node supports both text-to-video generation and image-to-video generation. The node will create corresponding payload depending on whether an input image is present or not."
-            ],
-            "color": "#432",
-            "bgcolor": "#000"
-        },
-        {
-            "id": 4,
-            "type": "LoadImage",
-            "pos": [
-                496.31859627609606,
-                -229.71277089860084
-            ],
-            "size": [
-                270,
-                314
-            ],
-            "flags": {},
-            "order": 1,
-            "mode": 0,
-            "inputs": [
-                {
-                    "localized_name": "image",
-                    "name": "image",
-                    "type": "COMBO",
-                    "widget": {
-                        "name": "image"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "choose file to upload",
-                    "name": "upload",
-                    "type": "IMAGEUPLOAD",
-                    "widget": {
-                        "name": "upload"
-                    },
-                    "link": null
-                }
-            ],
-            "outputs": [
-                {
-                    "localized_name": "IMAGE",
-                    "name": "IMAGE",
-                    "type": "IMAGE",
-                    "links": [
-                        25
-                    ]
-                },
-                {
-                    "localized_name": "MASK",
-                    "name": "MASK",
-                    "type": "MASK",
-                    "links": null
-                }
-            ],
-            "properties": {
-                "Node name for S&R": "LoadImage"
-            },
-            "widgets_values": [
-                "cute-cat.jpg",
-                "image"
-            ]
-        },
-        {
-            "id": 8,
-            "type": "VLLMOmniDiffusionSampling",
-            "pos": [
-                465.5140218220927,
-                148.80072646828987
-            ],
-            "size": [
-                284.205078125,
-                226
-            ],
-            "flags": {},
-            "order": 2,
-            "mode": 0,
-            "inputs": [
-                {
-                    "localized_name": "n",
-                    "name": "n",
-                    "type": "INT",
-                    "widget": {
-                        "name": "n"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "num_inference_steps",
-                    "name": "num_inference_steps",
-                    "type": "INT",
-                    "widget": {
-                        "name": "num_inference_steps"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "guidance_scale",
-                    "name": "guidance_scale",
-                    "type": "FLOAT",
-                    "widget": {
-                        "name": "guidance_scale"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "true_cfg_scale",
-                    "name": "true_cfg_scale",
-                    "type": "FLOAT",
-                    "widget": {
-                        "name": "true_cfg_scale"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "vae_use_slicing",
-                    "name": "vae_use_slicing",
-                    "type": "BOOLEAN",
-                    "widget": {
-                        "name": "vae_use_slicing"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "vae_use_tiling",
-                    "name": "vae_use_tiling",
-                    "type": "BOOLEAN",
-                    "widget": {
-                        "name": "vae_use_tiling"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "seed",
-                    "name": "seed",
-                    "type": "INT",
-                    "widget": {
-                        "name": "seed"
-                    },
-                    "link": null
-                }
-            ],
-            "outputs": [
-                {
-                    "localized_name": "diffusion sampling params",
-                    "name": "diffusion sampling params",
-                    "type": "SAMPLING_PARAMS",
-                    "links": [
-                        27
-                    ]
-                }
-            ],
-            "properties": {
-                "Node name for S&R": "VLLMOmniDiffusionSampling"
-            },
-            "widgets_values": [
-                1,
-                50,
-                1,
-                1,
-                false,
-                false,
-                1992,
-                "randomize"
-            ]
-        },
-        {
-            "id": 14,
-            "type": "VLLMOmniWanParams",
-            "pos": [
-                518.6005521845747,
-                428.4455663719664
-            ],
-            "size": [
-                270,
-                106
-            ],
-            "flags": {},
-            "order": 3,
-            "mode": 0,
-            "inputs": [
-                {
-                    "localized_name": "guidance_scale_2",
-                    "name": "guidance_scale_2",
-                    "type": "FLOAT",
-                    "widget": {
-                        "name": "guidance_scale_2"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "boundary_ratio",
-                    "name": "boundary_ratio",
-                    "type": "FLOAT",
-                    "widget": {
-                        "name": "boundary_ratio"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "flow_shift",
-                    "name": "flow_shift",
-                    "type": "FLOAT",
-                    "widget": {
-                        "name": "flow_shift"
-                    },
-                    "link": null
-                }
-            ],
-            "outputs": [
-                {
-                    "localized_name": "Wan video params",
-                    "name": "Wan video params",
-                    "type": "VIDEO_PARAMS",
-                    "links": [
-                        28
-                    ]
-                }
-            ],
-            "properties": {
-                "Node name for S&R": "VLLMOmniWanParams"
-            },
-            "widgets_values": [
-                4,
-                0.875,
-                5
-            ]
-        },
-        {
-            "id": 12,
-            "type": "VLLMOmniGenerateVideo",
-            "pos": [
-                827.2566336087867,
-                -73.77449831827545
-            ],
-            "size": [
-                400,
-                346
-            ],
-            "flags": {},
-            "order": 4,
-            "mode": 0,
-            "inputs": [
-                {
-                    "localized_name": "image",
-                    "name": "image",
-                    "shape": 7,
-                    "type": "IMAGE",
-                    "link": 25
-                },
-                {
-                    "localized_name": "sampling_params",
-                    "name": "sampling_params",
-                    "shape": 7,
-                    "type": "SAMPLING_PARAMS",
-                    "link": 27
-                },
-                {
-                    "localized_name": "lora",
-                    "name": "lora",
-                    "shape": 7,
-                    "type": "REMOTE_LORA",
-                    "link": null
-                },
-                {
-                    "localized_name": "model_params",
-                    "name": "model_params",
-                    "shape": 7,
-                    "type": "VIDEO_PARAMS",
-                    "link": 28
-                },
-                {
-                    "localized_name": "url",
-                    "name": "url",
-                    "type": "STRING",
-                    "widget": {
-                        "name": "url"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "model",
-                    "name": "model",
-                    "type": "STRING",
-                    "widget": {
-                        "name": "model"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "prompt",
-                    "name": "prompt",
-                    "type": "STRING",
-                    "widget": {
-                        "name": "prompt"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "negative_prompt",
-                    "name": "negative_prompt",
-                    "type": "STRING",
-                    "widget": {
-                        "name": "negative_prompt"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "width",
-                    "name": "width",
-                    "type": "INT",
-                    "widget": {
-                        "name": "width"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "height",
-                    "name": "height",
-                    "type": "INT",
-                    "widget": {
-                        "name": "height"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "fps",
-                    "name": "fps",
-                    "type": "INT",
-                    "widget": {
-                        "name": "fps"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "num_frames",
-                    "name": "num_frames",
-                    "type": "INT",
-                    "widget": {
-                        "name": "num_frames"
-                    },
-                    "link": null
-                }
-            ],
-            "outputs": [
-                {
-                    "localized_name": "video",
-                    "name": "video",
-                    "type": "VIDEO",
-                    "links": [
-                        26
-                    ]
-                }
-            ],
-            "properties": {
-                "Node name for S&R": "VLLMOmniGenerateVideo"
-            },
-            "widgets_values": [
-                "http://localhost:8000/v1",
-                "Wan-AI/Wan2.2-TI2V-5B-Diffusers",
-                "Make the cat in the attached video yawning. Eye closed and mouth wide open. Keep the original realistic style.",
-                "Cartoonish",
-                900,
-                1200,
-                12,
-                41
-            ]
-        },
-        {
-            "id": 13,
-            "type": "SaveVideo",
-            "pos": [
-                1257.5441369501657,
-                -44.12956394419863
-            ],
-            "size": [
-                374,
-                372.34896826553796
-            ],
-            "flags": {},
-            "order": 5,
-            "mode": 0,
-            "inputs": [
-                {
-                    "localized_name": "video",
-                    "name": "video",
-                    "type": "VIDEO",
-                    "link": 26
-                },
-                {
-                    "localized_name": "filename_prefix",
-                    "name": "filename_prefix",
-                    "type": "STRING",
-                    "widget": {
-                        "name": "filename_prefix"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "format",
-                    "name": "format",
-                    "type": "COMBO",
-                    "widget": {
-                        "name": "format"
-                    },
-                    "link": null
-                },
-                {
-                    "localized_name": "codec",
-                    "name": "codec",
-                    "type": "COMBO",
-                    "widget": {
-                        "name": "codec"
-                    },
-                    "link": null
-                }
-            ],
-            "outputs": [],
-            "properties": {},
-            "widgets_values": [
-                "video/ComfyUI",
-                "auto",
-                "auto"
-            ]
-        }
-    ],
-    "links": [
-        [
-            25,
-            4,
-            0,
-            12,
-            0,
-            "IMAGE"
-        ],
-        [
-            26,
-            12,
-            0,
-            13,
-            0,
-            "VIDEO"
-        ],
-        [
-            27,
-            8,
-            0,
-            12,
-            1,
-            "SAMPLING_PARAMS"
-        ],
-        [
-            28,
-            14,
-            0,
-            12,
-            3,
-            "VIDEO_PARAMS"
-        ]
-    ],
-    "groups": [
-        {
-            "id": 1,
-            "title": "Input",
-            "bounding": [
-                213.1706010087147,
-                -313.20095750667554,
-                560.8246471437027,
-                407.4246532472182
-            ],
-            "color": "#3f789e",
-            "font_size": 24,
-            "flags": {}
-        }
-    ],
-    "config": {},
-    "extra": {
-        "workflowRendererVersion": "LG",
-        "ds": {
-            "scale": 1.1469075819486894,
-            "offset": [
-                101.7632950847089,
-                492.29121889347806
-            ]
-        }
-    },
-    "version": 0.4
-}
+{"id":"3aea1d08-7de5-4997-ba71-b2f66ad72a21","revision":0,"last_node_id":24,"last_link_id":35,"nodes":[{"id":13,"type":"SaveVideo","pos":[1315.729153459605,-137.92038948394355],"size":[374,313.7692307692308],"flags":{},"order":9,"mode":0,"inputs":[{"localized_name":"video","name":"video","type":"VIDEO","link":32},{"localized_name":"filename_prefix","name":"filename_prefix","type":"STRING","widget":{"name":"filename_prefix"},"link":null},{"localized_name":"format","name":"format","type":"COMBO","widget":{"name":"format"},"link":null},{"localized_name":"codec","name":"codec","type":"COMFY_DYNAMICCOMBO_V3","widget":{"name":"codec"},"link":null}],"outputs":[{"localized_name":"video","name":"video","type":"VIDEO","links":null}],"properties":{},"widgets_values":["video/ComfyUI","auto","auto"]},{"id":10,"type":"MarkdownNote","pos":[183.74636797766794,-571.2529158046766],"size":[240.3326478922474,136.3505820791881],"flags":{},"order":0,"mode":0,"inputs":[],"outputs":[],"title":"Note: Task and Input","properties":{},"widgets_values":["vLLM-Omni nodes are categorized based on the output modality. The \"Generate Video\" node supports both *text-to-video*, *frame-to-video*, and *reference-to-video* generation. The node will create corresponding payload depending on whether an input image is present or not."],"color":"#432","bgcolor":"#000"},{"id":4,"type":"LoadImage","pos":[452.07189605801125,-569.7242536425791],"size":[270,314],"flags":{},"order":1,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[29]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["rap.png","image"]},{"id":23,"type":"LoadImage","pos":[9.206110246057657,-7.6955291509170785],"size":[270,314],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[34]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["white-cat.jpg","image"]},{"id":22,"type":"MarkdownNote","pos":[-9.783113411416666,-148.8790300038605],"size":[297.3012007349086,88],"flags":{},"order":3,"mode":0,"inputs":[],"outputs":[],"title":"Note: Reference to video (r2v)","properties":{},"widgets_values":["For r2v tasks, connect this Video Reference block to the main Generate Video block. Then disconnect the \"frame\" input above."],"color":"#432","bgcolor":"#000"},{"id":24,"type":"LoadAudio","pos":[317.9929742385805,-141.44821500065095],"size":[282.798828125,136],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"audio","name":"audio","type":"COMBO","widget":{"name":"audio"},"link":null},{"localized_name":"audioUI","name":"audioUI","type":"AUDIO_UI","widget":{"name":"audioUI"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"AUDIOUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"AUDIO","name":"AUDIO","type":"AUDIO","links":[35]}],"properties":{"Node name for S&R":"LoadAudio"},"widgets_values":["nothing-beats-a-jet2-holiday.mp3",null,""]},{"id":21,"type":"VLLMOmniVideoReferences","pos":[376.6129172897595,105.41557277556292],"size":[156.9380859375,126],"flags":{},"order":7,"mode":0,"inputs":[{"localized_name":"image_1","name":"image_1","shape":7,"type":"IMAGE","link":34},{"localized_name":"image_2","name":"image_2","shape":7,"type":"IMAGE","link":null},{"localized_name":"audio_1","name":"audio_1","shape":7,"type":"AUDIO","link":35},{"localized_name":"audio_2","name":"audio_2","shape":7,"type":"AUDIO","link":null},{"localized_name":"video_1","name":"video_1","shape":7,"type":"VIDEO","link":null},{"localized_name":"video_2","name":"video_2","shape":7,"type":"VIDEO","link":null}],"outputs":[{"localized_name":"references","name":"references","type":"VIDEO_REFERENCES","links":null}],"properties":{"Node name for S&R":"VLLMOmniVideoReferences"},"widgets_values":[]},{"id":8,"type":"VLLMOmniDiffusionSampling","pos":[83.24620208113915,422.08543477542605],"size":[284.205078125,226],"flags":{},"order":5,"mode":0,"inputs":[{"localized_name":"n","name":"n","type":"INT","widget":{"name":"n"},"link":null},{"localized_name":"num_inference_steps","name":"num_inference_steps","type":"INT","widget":{"name":"num_inference_steps"},"link":null},{"localized_name":"guidance_scale","name":"guidance_scale","type":"FLOAT","widget":{"name":"guidance_scale"},"link":null},{"localized_name":"true_cfg_scale","name":"true_cfg_scale","type":"FLOAT","widget":{"name":"true_cfg_scLine truncated

docker/Dockerfile.npu

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.25.0
+ARG VLLM_ASCEND_TAG=v0.26.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 # WORKDIR /vllm-workspace/vllm
@@ -26,6 +26,10 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
     export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
     python3 -m pip install -v -e /vllm-workspace/vllm-omni/ --no-build-isolation
 
+RUN python3 -m pip install -v --no-deps \
+      "stepaudio2-minicpmo>=0.1.1" \
+      "step-audio2>=1.0.0"
+
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 
 ENTRYPOINT []

docker/Dockerfile.npu.a3

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.25.0-a3
+ARG VLLM_ASCEND_TAG=v0.26.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 # WORKDIR /vllm-workspace/vllm
@@ -26,6 +26,10 @@ RUN export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi
     export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
     python3 -m pip install -v -e /vllm-workspace/vllm-omni/ --no-build-isolation
 
+RUN python3 -m pip install -v --no-deps \
+      "stepaudio2-minicpmo>=0.1.1" \
+      "step-audio2>=1.0.0"
+
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 
 ENTRYPOINT []

docker/Dockerfile.npu.ci

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.25.0
+ARG VLLM_ASCEND_TAG=v0.26.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -35,6 +35,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "psutil>=7.2.0" \
       "soundfile>=0.13.1" \
       "pyttsx3>=2.99" \
+      "opencc==1.4.1" \
       "opencv-python>=4.12.0.88" \
       "imageio[ffmpeg]>=0.6.0" \
       "jiwer>=3.0.0" \

docker/Dockerfile.npu.ci.a3

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.25.0-a3
+ARG VLLM_ASCEND_TAG=v0.26.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -35,6 +35,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "psutil>=7.2.0" \
       "soundfile>=0.13.1" \
       "pyttsx3>=2.99" \
+      "opencc==1.4.1" \
       "opencv-python>=4.12.0.88" \
       "imageio[ffmpeg]>=0.6.0" \
       "jiwer>=3.0.0" \
@@ -44,6 +45,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "hyperpyyaml>=1.2.0" \
       "onnx>=1.16.0" \
       "librosa>=0.10.0" \
+      "torchmetrics>=1.9.0" \
       --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
       --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
       --default-timeout=60 && \

docs/.nav.yml

@@ -118,6 +118,7 @@ nav:
       - design/feature/teacache.md
       - design/feature/skip_softmax.md
       - design/feature/async_chunk.md
+      - design/feature/omni_async_output_materialization.md
       - design/feature/vae_parallel.md
       - design/feature/diffusion_step_execution.md
       - design/feature/diffusion_request_level_batching.md

docs/README.md

@@ -52,12 +52,13 @@ vLLM-Omni is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
+- Full-duplex realtime serving with streaming audio input and output (experimental)
 
 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:
 
-- Omni-modality models (e.g. Qwen3-Omni, Cosmos3, HunyuanImage, BAGEL)
+- Omni-modality models (e.g. Qwen3-Omni, MiniCPM-o 4.5, Cosmos3, HunyuanImage, BAGEL)
 - TTS models (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
-- Diffusion models — image, video, and audio generation (e.g. Qwen-Image, Wan2.2, FLUX)
+- Diffusion models — image, video, and audio generation (e.g. MiniMax H3, Qwen-Image, Wan2.2, FLUX)
 - Robot-policy and action models (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)
 
 For more information, checkout the following:

docs/cli/README.md

@@ -18,18 +18,12 @@ Specify the port:
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091
 ```
 
-For a migrated model, load a custom deploy configuration with `--deploy-config`:
+Load a custom deploy configuration with `--deploy-config`:
 
 ```bash
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --deploy-config /path/to/deploy_config.yaml
 ```
 
-The deprecated `--stage-configs-path` flag is retained for models that still use the legacy `stage_args` schema:
-
-```bash
-vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --stage-configs-path /path/to/legacy_stage_config.yaml
-```
-
 ## bench
 
 Run benchmark tests for online serving throughput.

docs/cli/serve.md

@@ -8,7 +8,6 @@ The stage-based CLI is designed for deployments that require launching each pipe
 - For **migrated models** that utilize the bundled deployment YAML configurations located in
   `vllm_omni/deploy/`, the `--deploy-config` flag is only required to override the default configuration. By default, executing `vllm serve MODEL --omni ...`
   automatically loads the bundled deployment configuration.
-- For custom legacy `stage_args` YAMLs, pass the file with `--stage-configs-path`.
 
 Example: Initializing Stage 0 (Orchestrator and API Server):
 The commands below show a common device mapping where Stage 0 uses GPU 0 and
@@ -32,7 +31,7 @@ CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni \
     --omni-master-port 26000
 ```
 
-When utilizing a custom deployment YAML based on the new schema, append `--deploy-config /path/to/override.yaml` to each command execution. Conversely, for legacy models, substitute this parameter with `--stage-configs-path /path/to/stage_configs.yaml`.
+When utilizing a custom deployment YAML, append `--deploy-config /path/to/override.yaml` to each command execution.
 
 In the standard execution paradigm, the `--stage-overrides` argument is utilized to apply stage-specific configurations from a single CLI command.
 However, under the **stage-based CLI** paradigm, where each process strictly encapsulates a single stage, it is recommended to specify tuning parameters directly via discrete command-line flags for the respective stage, rather than constructing a composite `--stage-overrides` JSON string.

docs/configuration/README.md

@@ -2,7 +2,7 @@
 
 This section lists the most common options for running vLLM-Omni.
 
-For options within a vLLM Engine. Please refer to [vLLM Configuration](https://docs.vllm.ai/en/v0.16.0/configuration/index.html)
+For options within a vLLM Engine, please refer to the [vLLM 0.26 configuration guide](https://docs.vllm.ai/en/v0.26.0/configuration/index.html).
 
 Currently, the main options are maintained by stage configs for each model.
 

docs/configuration/composable_parallel.md

@@ -34,12 +34,8 @@ Two additional `StrategySpec` slots — `layer_hook_specs` and `kernel_specs`
 | 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`
@@ -216,17 +212,6 @@ Strategy keys MUST be `model_stage` *names* (strings), the same labels stage con
 !!! 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 (`--tensor-parallel-size 2`, `--stage-overrides '{"0": {"tensor_parallel_size": 2}}'`) 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:
 

docs/configuration/pd_disaggregation.md

@@ -158,7 +158,7 @@ runtime:
 
 ```bash
 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091 \
-    --stage-configs-path /path/to/qwen3_omni_pd.yaml
+    --deploy-config /path/to/qwen3_omni_pd.yaml
 ```
 
 ## Operational Notes

docs/configuration/stage_configs.md

@@ -1,9 +1,9 @@
-# Deploy and legacy stage configurations
+# Pipeline and deploy configurations
 
-In vLLM-Omni, a model's `PipelineConfig` defines its fixed stage topology, while a deploy configuration controls how those stages run. Models that have not migrated to this split still use legacy stage configurations with a `stage_args` schema.
+In vLLM-Omni, a model's `PipelineConfig` defines its fixed stage topology, while a deploy configuration controls how those stages run.
 
 !!! note
-    Default deploy config YAMLs (for example, `vllm_omni/deploy/qwen2_5_omni.yaml`, `vllm_omni/deploy/qwen3_omni_moe.yaml`, and `vllm_omni/deploy/qwen3_tts.yaml`) are bundled and loaded automatically when neither `--stage-configs-path` nor `--deploy-config` is provided — the model registry resolves the right pipeline + deploy YAML by `model_type`. Custom legacy `stage_args` YAMLs are still accepted via `--stage-configs-path`.
+    Default deploy config YAMLs (for example, `vllm_omni/deploy/qwen2_5_omni.yaml`, `vllm_omni/deploy/qwen3_omni_moe.yaml`, and `vllm_omni/deploy/qwen3_tts.yaml`) are bundled and loaded automatically when `--deploy-config` is omitted. The model registry resolves the right pipeline and deploy YAML by `model_type`.
 
 ## Deploy configuration schema
 
@@ -86,10 +86,9 @@ stages:
 
 | Flag | Description |
 |------|-------------|
-| `--deploy-config PATH` | Load a new-schema deploy YAML. It is mutually exclusive with `--stage-configs-path`. **Optional** — when omitted, the bundled `vllm_omni/deploy/<model_type>.yaml` is auto-loaded by the model registry. |
+| `--deploy-config PATH` | Load a deploy YAML. **Optional** — when omitted, the bundled `vllm_omni/deploy/<model_type>.yaml` is auto-loaded by the model registry. |
 | `--stage-overrides JSON` | Per-stage JSON overrides, e.g. `'{"0":{"gpu_memory_utilization":0.5}}'`. Per-stage values always win over global flags. |
 | `--async-chunk` / `--no-async-chunk` | Flip the deploy YAML's `async_chunk:` bool. Unset (default) leaves the YAML value in force. |
-| `--stage-configs-path` | **Deprecated.** Load a legacy `stage_args` YAML. New-schema YAMLs passed to this flag are auto-detected for compatibility, but should use `--deploy-config`. It is mutually exclusive with `--deploy-config` and will be removed in a future release. |
 
 ### Stage-Based CLI Paradigm
 
@@ -119,8 +118,7 @@ CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni \
     --omni-master-port 26000
 ```
 
-When instantiating a custom deployment YAML conforming to the updated schema, append the `--deploy-config /path/to/override.yaml` directive
-to all node invocations. For legacy architectures (e.g., BAGEL) configured via deprecated `stage_args:` schemas, continue to specify the relevant configuration via `--stage-configs-path /path/to/config.yaml`.
+When instantiating a custom deployment YAML, append the `--deploy-config /path/to/override.yaml` directive to all node invocations.
 
 In the context of standard initialization architectures, utilizing the `--stage-overrides` parameter operates as the optimal methodology
 for delineating stage-specific tuning from the CLI interface:
@@ -220,8 +218,7 @@ Therefore, as a core part of vLLM-Omni, a model's pipeline and deployment config
 
 To override specific parameters, explicitly inject the customized configuration schema
 in both online and offline instantiation flows. Use the `--deploy-config` flag
-when loading a deploy configuration, reserving the `--stage-configs-path` parameter
-exclusively to maintain compatibility with legacy `stage_args` YAML constructs.
+when loading a deploy configuration.
 
 Examples:
 
@@ -236,11 +233,6 @@ For online serving:
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --deploy-config /path/to/deploy_config.yaml
 ```
 
-Legacy online serving:
-
-```bash
-vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
-```
 !!! important
     We are actively iterating on the definition of deployment configurations, and we welcome feedback from users and developers.
 

docs/contributing/ci/test_system_overview.md

@@ -56,7 +56,11 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
       <td>/</td>
       <td><code>core_model and cpu</code></td>
       <td rowspan="2">&lt;15min</td>
-      <td>/tests/{component_name}/test_xxx</td>
+      <td>
+        <code>tests/{component}/…</code> mirroring <code>vllm_omni/{component}/</code><br>
+        (e.g. <code>tests/diffusion/</code>, <code>tests/engine/</code>, <code>tests/entrypoints/</code>)<br>
+        Do <strong>not</strong> add new top-level <code>tests/</code> dirs unrelated to a component.
+      </td>
       <td>
         <a href="test_writing_guide.md#l1--l2-level-testing-unit-testing-and-basic-end-to-end-verification">L1 &amp; L2</a><br>
         Section 1 L1&amp;L2: Purpose, Test Content, Directory Location, Example
@@ -68,16 +72,18 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
       <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, non-empty output, format match (no Whisper/accuracy)</td>
-      <td>High-priority features (using random lightweight models)</td>
+      <td>High-priority features and component tests that require launching instances</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>
       <td>
-        <strong>Model tests:</strong><br>
-        /tests/e2e/online_serving/test_{model_name}.py<br>
-        <strong>Feature tests:</strong><br>
-        /tests/{component_name}/test_xxx<br>
-        <strong>Interface tests:</strong><br>
-        /tests/entrypoints/test_xxx
+        <strong>Model E2E:</strong><br>
+        <code>tests/e2e/online_serving/test_{model}.py</code><br>
+        <code>tests/e2e/offline_inference/test_{model}.py</code><br>
+        <strong>Feature integration:</strong><br>
+        <code>tests/e2e/features/&lt;feature&gt;/</code><br>
+        (e.g. <code>fullduplex/</code>, <code>custom_pipeline/</code>, <code>rlhf_test/</code>)<br>
+        <strong>Component / interface Test (GPU):</strong><br>
+        <code>tests/{component}/…</code>, <code>tests/entrypoints/…</code>
       </td>
       <td>
         <a href="test_writing_guide.md#l1--l2-level-testing-unit-testing-and-basic-end-to-end-verification">L1 &amp; L2</a><br>
@@ -90,18 +96,19 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
       <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 + 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 features and component tests that require launching instances</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>
       <td>&lt;30min</td>
       <td>
-        <strong>Model tests:</strong><br>
-        /tests/e2e/online_serving/test_{model_name}.py<br>
-        /tests/e2e/offline_inference/test_{model_name}.py<br>
-        <strong>Feature tests:</strong><br>
-        /tests/{component_name}/test_xxx<br>
-        <strong>Interface tests:</strong><br>
-        /tests/entrypoints/test_xxx
+        <strong>Model E2E:</strong><br>
+        <code>tests/e2e/online_serving/test_{model}.py</code><br>
+        <code>tests/e2e/offline_inference/test_{model}.py</code><br>
+        <code>tests/e2e/accuracy/</code><br>
+        <strong>Feature integration:</strong><br>
+        <code>tests/e2e/features/&lt;feature&gt;/</code><br>
+        <strong>Component / interface Test:</strong><br>
+        <code>tests/{component}/…</code>, <code>tests/entrypoints/…</code>
       </td>
       <td>
         <a href="test_writing_guide.md#l3-level-testing-core-integration-performance-and-accuracy-verification">L3</a><br>
@@ -114,26 +121,23 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
       <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-priority models: function and doc testing</td>
-      <td>Low-priority features (using real weights)</td>
+      <td>Low-priority features and component tests that require launching instances</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>
-        <strong>Model tests:</strong><br>
-        /tests/e2e/online_serving/test_{model_name}_expansion.py<br>
-        <strong>Feature tests:</strong><br>
-        /tests/{component_name}/test_xxx<br>
-        <strong>Interface tests:</strong><br>
-        /tests/entrypoints/test_xxx<br>
+        <strong>Model E2E:</strong><br>
+        <code>tests/e2e/online_serving/test_{model}_expansion.py</code><br>
+        <code>tests/e2e/offline_inference/test_{model}_expansion.py</code><br>
+        <code>tests/e2e/accuracy/test_{model}.py</code><br>
+        <strong>Feature integration:</strong><br>
+        <code>tests/e2e/features/&lt;feature&gt;/</code><br>
+        <strong>Component / interface Test:</strong><br>
+        <code>tests/{component}/…</code>, <code>tests/entrypoints/…</code><br>
         <strong>Performance:</strong><br>
-        /tests/dfx/perf/tests/test_qwen3_omni_*.json (Omni), test_tts.json (TTS),<br>
-        test_voxcpm2.json, test_higgs_audio_v3.json, and<br>
-        /tests/dfx/perf/tests/test_{diffusion_model}_vllm_omni.json (Diffusion)<br>
-        <strong>Doc Test:</strong><br>
-        tests/examples/online_serving/test_{model_name}.py<br>
-        tests/examples/offline_inference/test_{model_name}.py<br>
-        <strong>Accuracy Test:</strong><br>
-        /tests/e2e/accuracy/test_{model_name}.py
+        <code>tests/dfx/perf/tests/</code><br>
+        <strong>Doc examples:</strong><br>
+        <code>tests/examples/online_serving/</code>, <code>tests/examples/offline_inference/</code>
       </td>
       <td>
         <a href="test_writing_guide.md#l4-level-testing-full-functionality-performance-and-documentation-testing">L4</a><br>
@@ -152,11 +156,9 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
       <td> Depends on reality </td>
       <td>
         <strong>Stability:</strong><br>
-        /tests/dfx/stability/tests/test_qwen3_omni.json<br>
-        /tests/dfx/stability/tests/test_wan22.json<br>
+        <code>tests/dfx/stability/tests/</code><br>
         <strong>Reliability:</strong><br>
-        tests/dfx/reliability/test_reliability_{model_key}.py<br>
-        (e.g. <code>test_reliability_qwen3_omni.py</code>, <code>test_reliability_wan22.py</code>, <code>test_reliability_hunyuan_image.py</code>, <code>test_reliability_voxcpm2.py</code>)
+        <code>tests/dfx/reliability/test_reliability_{model}.py</code>
       </td>
       <td>
         <a href="test_writing_guide.md#l5-level-testing-stability-and-reliability-testing">L5</a><br>
@@ -168,7 +170,9 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
   </tbody>
 </table>
 
-For per-level test authoring (directories, markers, examples), see [Test Writing Guide](./test_writing_guide.md).
+**Test Dir placement (summary):** component / unit under `tests/{component}/` mirroring `vllm_omni/`; model E2E under `tests/e2e/online_serving/`, `tests/e2e/offline_inference/`, and `tests/e2e/accuracy/`; feature integration under `tests/e2e/features/<feature>/`; doc example tests under `tests/examples/online_serving/` and `tests/examples/offline_inference/`; performance under `tests/dfx/perf/`; stability under `tests/dfx/stability/`; reliability under `tests/dfx/reliability/`. Do **not** add new top-level directories under `tests/` that are unrelated to a `vllm_omni` component (or to the established `e2e` / `dfx` / `helpers` / `examples` / `buildkite` layout).
+
+For per-level test authoring (markers, examples), see [Test Writing Guide](./test_writing_guide.md).
 
 ## Common Specifications
 

docs/contributing/ci/test_writing_guide.md

@@ -4,137 +4,20 @@ This document describes how to author tests for each CI level (L1–L5). For the
 
 ## Test Directory Structure
 
-Legend: `✅` = test exists, `⬜` = suggested to add.
-```
-vllm_omni/                                    tests/
-├── config/                             →     ├── config/
-│   ├── model.py                              │   └── test_model.py                    ⬜
-│   └── lora.py                               │   └── test_lora.py                      ⬜
-│
-├── core/                               →     ├── core/
-│   └── sched/                                 │   └── sched/
-│       ├── omni_ar_scheduler.py               │       ├── test_omni_ar_scheduler.py    ⬜
-│       ├── omni_generation_scheduler.py       │       ├── test_omni_generation_scheduler.py  ⬜
-│       └── output.py                          │       └── test_output.py               ✅ currently in entrypoints/test_omni_new_request_data.py (tests output.OmniNewRequestData)
-│
-├── diffusion/                          →     ├── diffusion/
-│   ├── diffusion_engine.py                    │   ├── test_diffusion_engine.py          ⬜
-│   ├── attention/                             │   ├── attention/
-│   │   ├── layer.py                            │   │   ├── test_attention_sp.py         ✅
-│   │   └── backends/                           │   │   └── test_flash_attn.py           ✅
-│   ├── distributed/                           │   ├── distributed/
-│   │   └── ...                                 │   │   ├── test_comm.py                 ✅
-│   │                                            │   │   ├── test_cfg_parallel.py        ✅
-│   │                                            │   │   └── test_sp_plan_hooks.py       ✅
-│   ├── lora/                                   │   ├── lora/
-│   │   └── ...                                 │   │   ├── test_base_linear.py          ✅
-│   │                                            │   │   └── test_lora_manager.py        ✅
-│   ├── models/                                 │   ├── models/
-│   │   ├── qwen_image/                         │   │   ├── qwen_image/                 (e2e coverage)
-│   │   ├── ovis_image/                         │   │   ├── ovis_image/
-│   │   │   └── ...                             │   │   │   └── test_ovis_image.py     ✅
-│   │   ├── z_image/                            │   │   └── z_image/
-│   │   └── ...                                 │   │       └── test_zimage_tp_constraints.py  ✅
-│   └── worker/                                 │   └── worker/
-│       ├── diffusion_worker.py                 │       └── test_diffusion_worker.py   ✅ file at diffusion/test_diffusion_worker.py
-│       └── diffusion_model_runner.py            │
-│
-├── distributed/                         →     ├── distributed/
-│   └── omni_connectors/                         │   └── omni_connectors/
-│       ├── adapter.py                           │       ├── test_adapter_and_flow.py   ✅
-│       ├── kv_transfer_manager.py               │       ├── test_basic_connectors.py   ✅
-│       ├── connectors/                           │       ├── test_kv_flow.py             ✅
-│       └── utils/                               │       └── test_omni_connector_configs.py  ✅
-│
-├── engine/                             →     ├── engine/
-│   ├── input_processor.py                      │   ├── test_input_processor.py         ⬜  (no processor.py in source)
-│   ├── output_processor.py                     │   └── test_output_processor.py         ⬜
-│   └── arg_utils.py                            │   └── test_arg_utils.py               ⬜
-│
-├── entrypoints/                        →     ├── entrypoints/
-│   ├── stage_utils.py                          │   ├── test_stage_utils.py            ✅
-│   ├── cli/                                     │   ├── cli/                           (benchmarks/test_serve_cli.py covers CLI serve)
-│   │   └── ...                                  │   │   └── test_*.py                  ⬜
-│   └── openai/                                  │   └── openai_api/                    # maps to entrypoints/openai/
-│       ├── api_server.py                        │       ├── test_api_server.py         ⬜  (e2e indirect coverage)
-│       ├── serving_chat.py                       │       ├── test_serving_chat_sampling_params.py  ✅
-│       ├── serving_speech.py                     │       ├── test_serving_speech.py     ✅
-│       └── image_api_utils.py                   │       └── test_image_server.py      ✅
-│
-├── inputs/                             →     ├── inputs/
-│   ├── data.py                                 │   ├── test_data.py                   ⬜
-│   ├── parse.py                                │   ├── test_parse.py                 ⬜
-│   └── preprocess.py                            │   └── test_preprocess.py            ✅ currently in entrypoints/test_omni_input_preprocessor.py
-│
-├── model_executor/                     →     ├── model_executor/
-│   ├── layers/                                  │   ├── layers/
-│   │   └── mrope.py                             │   │   └── test_mrope.py              ⬜
-│   ├── model_loader/                            │   ├── model_loader/
-│   │   └── weight_utils.py                      │   │   └── test_weight_utils.py      ⬜
-│   ├── models/                                  │   ├── models/
-│   │   ├── qwen2_5_omni/                         │   │   ├── qwen2_5_omni/
-│   │   │   ├── qwen2_5_omni_thinker.py           │   │   │   ├── test_audio_length.py  ✅
-│   │   │   ├── qwen2_5_omni_talker.py            │   │   │   ├── test_qwen2_5_omni_thinker.py  ⬜
-│   │   │   └── qwen2_5_omni_token2wav.py         │   │   │   ├── test_qwen2_5_omni_talker.py  ⬜
-│   │   └── qwen3_omni/                          │   │   │   └── test_qwen2_5_omni_token2wav.py  ⬜
-│   │       └── ...                               │   │   └── qwen3_omni/
-│   ├── stage_configs/                           │   │       └── test_*.py              ⬜
-│   │   └── *.yaml                               │   └── stage_configs/                 (used by e2e, test_*.py can be added)  ⬜
-│   └── stage_input_processors/                  │   └── stage_input_processors/
-│       └── ...                                  │       └── test_*.py                 ⬜
-│
-├── sample/                             →     ├── sample/
-│   └── __init__.py                             │   └── test_*.py                      ⬜
-│
-├── utils/                              →     ├── utils/
-│   └── __init__.py                             │   └── test_*.py                       ⬜  (no platform_utils.py currently)
-│
-├── worker/                             →     ├── worker/
-│   ├── gpu_ar_model_runner.py                  │   ├── test_gpu_ar_model_runner.py    ⬜
-│   ├── gpu_ar_worker.py                        │   ├── test_gpu_ar_worker.py           ⬜
-│   ├── gpu_generation_model_runner.py          │   ├── test_gpu_generation_model_runner.py  ✅
-│   ├── gpu_generation_worker.py                │   ├── test_gpu_generation_worker.py  ⬜
-│   ├── gpu_model_runner.py                     │   ├── test_omni_gpu_model_runner.py   ✅
-│   └── mixins.py                               │   └── (npu under platforms/npu/worker/)  # not worker/npu/
-│
-├── platforms/                          →     (no tests/platforms/, e2e and stage_configs provide indirect coverage)
-│   ├── cuda/
-│   ├── npu/worker/                             # NPU worker here, not vllm_omni/worker/npu/
-│   ├── rocm/
-│   └── xpu/worker/
-│
-├── outputs.py                          →     test_outputs.py                         ✅ (at tests root)
-├── (logger, patch, request, version)    →     (no corresponding unit test)
-│
-└── e2e (tests side only)               →     ├── e2e/
-                                               ├── online_serving/                     ✅ non-empty
-                                               │   ├── test_qwen2_5_omni.py
-                                               │   ├── test_async_omni.py
-                                               │   ├── test_qwen3_omni.py
-                                               │   ├── test_qwen3_omni_expansion.py
-                                               │   ├── test_mimo_audio.py
-                                               │   └── test_images_generations_lora.py
-                                               └── offline_inference/                  ✅
-                                                   ├── test_qwen2_5_omni.py
-                                                   ├── test_qwen3_omni.py
-                                                   ├── test_bagel_text2img.py
-                                                   ├── test_z_image.py
-                                                   ├── test_wan22.py
-                                                   ├── test_zimage_tensor_parallel.py
-                                                   ├── test_cache_dit.py
-                                                   ├── test_teacache.py
-                                                   ├── test_stable_audio_expansion.py
-                                                   ├── test_diffusion_cpu_offload.py
-                                                   ├── test_diffusion_layerwise_offload.py
-                                                   ├── test_diffusion_lora.py
-                                                   ├── test_sequence_parallel.py
-                                                   └── stage_configs/                  (legacy schema, still
-                                                       ├── bagel_*.yaml                 present for unmigrated
-                                                       └── npu/, rocm/, etc.            models)
+Canonical placement by CI level is in the **Test Dir** column of the level matrix in [Test System Overview](./test_system_overview.md).
 
-# Migrated models (qwen3_omni_moe, qwen2_5_omni, qwen3_tts) live under
-# vllm_omni/deploy/ instead — see docs/configuration/stage_configs.md.
-```
+Summary:
+
+- **Component / unit:** `tests/{component}/…` mirroring `vllm_omni/{component}/`
+- **Model E2E:** `tests/e2e/online_serving/`, `tests/e2e/offline_inference/`, `tests/e2e/accuracy/`
+- **Feature integration:** `tests/e2e/features/<feature>/`
+- **Doc example tests:** `tests/examples/online_serving/`, `tests/examples/offline_inference/`
+- **Performance:** `tests/dfx/perf/tests/` (JSON configs; runners under `tests/dfx/perf/scripts/`)
+- **Stability:** `tests/dfx/stability/tests/`
+- **Reliability:** `tests/dfx/reliability/`
+- Do **not** add new top-level `tests/` directories unrelated to a component (or to `e2e` / `dfx` / `helpers` / `examples` / `buildkite`)
+
+Deploy YAMLs: `vllm_omni/deploy/` via `tests.helpers.stage_config.get_deploy_config_path` (see [stage configs](../../configuration/stage_configs.md)).
 
 ## Markers for Tests
 
@@ -373,13 +256,13 @@ L2 level testing builds upon L1 by introducing GPU resources and verifying that
 
 #### 1.3 Test Directory and Execution Files
 
-A clear directory structure is key to managing test cases efficiently.
+A clear directory structure is key to managing test cases efficiently. See the **Test Dir** column in [Test System Overview](./test_system_overview.md) for the canonical placement rules.
 
--   ***L1 Test Directory***: `/tests/{component_name}/test_xxx.py`
--   -   Here, `{component_name}` corresponds to modules in the source code, such as `distributed`, `entrypoints`, etc., and `test_xxx.py` is the specific test file.
--   ***L2 Test Directory***:
--   -   Online Serving: `/tests/e2e/online_serving/test_{model_name}.py`
-    -   Offline Inference: `/tests/e2e/offline_inference/test_{model_name}.py`
+-   ***L1 (component / unit)***: `tests/{component_name}/…` mirroring `vllm_omni/{component_name}/` (e.g. `tests/diffusion/`, `tests/entrypoints/`). Do **not** invent new top-level `tests/` folders unrelated to a component.
+-   ***L2 (basic model E2E)***:
+    -   Online serving: `tests/e2e/online_serving/test_{model_name}.py`
+    -   Offline inference: `tests/e2e/offline_inference/test_{model_name}.py`
+-   ***Feature-level integration*** (cross-cutting, not a single-model smoke): `tests/e2e/features/<feature>/` (e.g. `fullduplex/`, `custom_pipeline/`, `rlhf_test/`).
 
 #### 1.4 Execution Method and Example
 
@@ -485,10 +368,12 @@ L3 level testing executes after code is merged into the main branch. Its core pu
 
 #### 2.3 Test Directory and Execution Files
 
--   ***Functional Testing***:
--   -   Online Serving: `/tests/e2e/online_serving/test_{model_name}_expansion.py`
-    -   Offline Inference: `/tests/e2e/offline_inference/test_{model_name}_expansion.py`
-    -   (Note: `_expansion.py` likely means it contains more comprehensive scenario cases compared to L2 tests).
+-   ***Functional / model E2E***:
+    -   Online serving: `tests/e2e/online_serving/test_{model_name}_expansion.py`
+    -   Offline inference: `tests/e2e/offline_inference/test_{model_name}_expansion.py`
+    -   Accuracy / similarity: `tests/e2e/accuracy/`
+    -   (`_expansion.py` holds broader scenario coverage than the L2 smoke file.)
+-   ***Feature-level integration***: `tests/e2e/features/<feature>/` (not under a new top-level `tests/` directory).
 
 #### 2.4 Execution Method and Example
 

docs/contributing/model/adding_diffusion_model.md

@@ -653,7 +653,7 @@ python end2end.py --model your-org/your-model-name --modality text2img --prompts
 Mirror BAGEL’s online serving setup:
 
 - Server launcher: `examples/online_serving/your_model_name/run_server.sh`
-  - Wrap `vllm serve ... --omni --port ...` (and `--stage-configs-path ...` if needed)
+  - Wrap `vllm serve ... --omni --port ...` (and `--deploy-config ...` if needed)
 - Client: `examples/online_serving/your_model_name/openai_chat_client.py`
   - Send requests to `POST /v1/chat/completions`
   - Support multimodal inputs (e.g., base64 image) if your model needs it

docs/contributing/model/adding_omni_model.md

@@ -48,8 +48,8 @@ vllm_omni/deploy/
 ```
 
 New in-tree models should register their `PipelineConfig` in
-`vllm_omni/config/pipeline_registry.py`; use `--stage-configs-path` only for
-custom legacy `stage_args` YAMLs.
+`vllm_omni/config/pipeline_registry.py`; use `--deploy-config` for custom
+deployment YAMLs.
 
 ## Step-by-Step Implementation
 

docs/contributing/model/adding_tts_model.md

@@ -449,9 +449,8 @@ Key configuration fields:
 !!! note
     New in-tree models should define frozen topology in `models/<model>/pipeline.py`,
     register it in `vllm_omni/config/pipeline_registry.py`, and put deployment
-    knobs in `vllm_omni/deploy/<model>.yaml`. Legacy `stage_args` YAMLs are still
-    accepted for custom configs via `--stage-configs-path`, but should not be used
-    for new bundled defaults.
+    knobs in `vllm_omni/deploy/<model>.yaml`. Use `--deploy-config` to load a
+    custom deployment YAML.
 
 ### Batch mode
 

docs/design/feature/async_diffusion_output.md

@@ -0,0 +1,174 @@
+# Async Diffusion Output
+
+## Table of Contents
+
+1. [Overview](#overview)
+2. [Performance](#performance)
+3. [Architecture](#architecture)
+4. [Model Coverage](#model-coverage)
+5. [Configuration](#configuration)
+6. [Related Files](#related-files)
+
+## Overview
+
+The async diffusion output feature moves the D2H (Device→Host) copy and SHM packing from the Worker main thread to a background daemon thread, so the GPU default stream can immediately start the next request's forward pass. This eliminates the GPU bubble caused by synchronous D2H/packing.
+
+**Core benefit**: the previous request's output D2H/packing overlaps with the next request's forward pass.
+
+**Automatically enabled** when `step_execution=False` (default, request-mode). No extra configuration needed. When `step_execution=True` (step-mode), the original synchronous path is used — pump thread and Worker background thread are not started.
+
+## Performance
+
+### HunyuanImage-3.0 (TP4)
+
+| Resolution | Async Off (QPS) | Async On (QPS) | Change |
+|---|---|---|---|
+| 1024×1024 | 0.4773 | 0.4802 | +0.60% |
+| 768×768 | 0.8370 | 0.8533 | +1.95% |
+
+**Analysis**: smaller resolution → larger QPS gain. D2H/packing is relatively fixed overhead; forward is variable (scales with resolution). Smaller images → shorter forward → D2H/packing occupies a larger fraction of total time → more benefit from overlapping it.
+
+## Architecture
+
+### Execution Timeline
+
+```
+Before (synchronous D2H):
+[forward req1] [D2H+SHM pack req1] [forward req2] [D2H+SHM pack req2]
+                                 ^^^^^^^^^^^^^^^^ GPU bubble
+
+After (async D2H):
+[forward req1] [forward req2]
+[D2H+SHM pack req1 (side stream)] [D2H+SHM pack req2 (side stream)]
+                                 ^^^^^^^^^^^^^^^^ bubble eliminated
+```
+
+### Data Flow
+
+```
+                    ┌─────────────────────────────────────────────────────┐
+                    │                 Worker Process                      │
+                    │                                                    │
+  execute_model ──> │ WorkerBusyLoop ──> forward() ──> DiffusionOutput   │
+                    │       │                              │ (GPU tensor)│
+                    │       │                    ┌─────────┴──────────┐  │
+                    │       │                    │  async output path │  │
+                    │       │                    │                    │  │
+                    │       │              compute_done ──> result_mq │  │
+                    │       │                    │                    │  │
+                    │       │                    ▼                    │  │
+                    │       │            AsyncOutputThread            │  │
+                    │       │         (side CUDA stream)             │  │
+                    │       │         wait_event(gpu_event)          │  │
+                    │       │         pack_diffusion_output_shm()    │  │
+                    │       │         D2H + SHM write                │  │
+                    │       │                    │                    │  │
+                    │       │              output_ready ──> result_mq │  │
+                    │       │                    │                    │  │
+                    │       ▼                    ▼                    │  │
+                    │  (dequeue next request)                        │  │
+                    └─────────────────────────────────────────────────────┘
+                                         │
+                                         ▼
+                    ┌─────────────────────────────────────────────────────┐
+                    │              Executor (main process)               │
+                    │                                                    │
+                    │  ResultPumpThread (sole reader of result_mq)       │
+                    │       │                                            │
+                    │       ├── AsyncDiffusionOutput(COMPUTE_DONE)       │
+                    │       │   → resolve _rpc_futures[rpc_id]           │
+                    │       │   → RunnerOutput(async_output_id=...)      │
+                    │       │                                            │
+                    │       ├── AsyncDiffusionOutput(OUTPUT_READY)       │
+                    │       │   → resolve _output_futures[async_output_id]│
+                    │       │   → or batch split via _batch_split_map   │
+                    │       │                                            │
+                    │       └── Non-async message                        │
+                    │           → _sync_result_buffer (for other RPCs)   │
+                    │                                                    │
+                    │  wait_output_ready(async_output_id) → Future       │
+                    └─────────────────────────────────────────────────────┘
+```
+
+### Key Components
+
+1. **AsyncDiffusionOutput** (`data.py`): Protocol envelope for `result_mq`. `kind` field routes messages:
+   - `COMPUTE_DONE` — forward finished, GPU can start next request
+   - `OUTPUT_READY` — D2H/SHM packing finished, final output available
+   - `RPC_RESULT` — ordinary RPC return (including error propagation)
+
+2. **Worker AsyncOutputThread** (`diffusion_worker.py`): Background daemon thread that:
+   - Waits for `gpu_event` (cross-stream sync ensuring forward wrote the tensor)
+   - Runs `pack_diffusion_output_shm()` on a side CUDA stream
+   - Enqueues `OUTPUT_READY` to `result_mq` when packing completes
+
+3. **ResultPumpThread** (`multiproc_executor.py`): Sole reader of `result_mq` in request-mode. Dispatches messages:
+   - `COMPUTE_DONE` / `RPC_RESULT` → resolves `_rpc_futures[rpc_id]`
+   - `OUTPUT_READY` → resolves `_output_futures[async_output_id]` (with batch split via `_batch_split_map`)
+   - Non-async messages → `_sync_result_buffer` for other RPCs
+
+4. **collective_rpc() two-path dispatch** (`multiproc_executor.py`):
+   - **Path 1**: `execute_model` / `execute_model_batch` → generates `rpc_id`, registers Future, waits for pump to deliver `compute_done`
+   - **Path 2**: All other RPCs → `_sync_result_buffer` (pump-fed) or `_result_mq` (step-mode, no pump)
+
+5. **record_device_event** (`platforms/`): Cross-platform GPU event recording for side-stream synchronization. Implemented for CUDA, NPU, ROCm, XPU, MUSA; base class returns `None` (safe no-op).
+
+6. **IPC side-stream D2H** (`ipc.py`): `pack_diffusion_output_shm()` and helpers accept `d2h_stream` parameter; use `pin_memory` + `copy_(non_blocking=True)` on the side stream instead of synchronous `.cpu()`.
+
+### Batch Split
+
+When `execute_model_batch` returns a `COMPUTE_DONE` with a single `async_output_id` for the entire batch, the executor splits it into per-request `async_output_id`s (format `{batch_id}/{request_id}`) via `_batch_split_map`. When `OUTPUT_READY` arrives, the pump extracts per-request results from the batch output and resolves each request's output future independently.
+
+## Model Coverage
+
+### Currently Supported (request-mode, `step_execution=False`)
+
+All models running in request-mode (`step_execution=False`, the default) automatically use async output.
+
+**Verified models**:
+
+| Model | Type | `supports_request_batch` |
+|---|---|---|
+| **HunyuanImage-3.0** | Image | `False` |
+| **Qwen-Image** | Image | `True` |
+
+Other models with `step_execution=False` are also supported but not yet verified.
+
+### Not Yet Supported (step-mode, `step_execution=True`)
+
+When `step_execution=True` (or `streaming_output=True`, which auto-enables step-mode), models use `execute_stepwise` instead of `execute_model`, which is not in the async Path 1 whitelist. Async output is not applicable.
+
+**Verified**: Helios (`step_execution=True`, does not benefit from this feature).
+
+Other models with `step_execution=True` are also not applicable.
+
+Adapting step-mode requires additional design because:
+
+- Each step's `RunnerOutput` must be available synchronously for the scheduler to advance
+- Intermediate latent tensors may not need D2H (only the final step needs D2H + postprocess)
+- Steps are serial (step N+1 depends on step N output), so the overlap benefit is limited to "final step's D2H overlaps with the next request's forward"
+
+**Planned approach**: first cover only the final step's async D2H, keep intermediate steps synchronous. This matches the request-mode benefit pattern.
+
+## Configuration
+
+No configuration needed. Async output is automatically enabled when `step_execution=False` (default).
+
+| `step_execution` | Mode | Async Output | Pump Thread | Worker BG Thread |
+|---|---|---|---|---|
+| `False` (default) | request-mode | ✅ enabled | ✅ started | ✅ started |
+| `True` | step-mode | ❌ disabled | ❌ not started | ❌ not started |
+
+## Related Files
+
+- `vllm_omni/diffusion/data.py`: `AsyncDiffusionOutput`, `AsyncOutputKind`, `DiffusionOutput.async_output_id`
+- `vllm_omni/diffusion/worker/diffusion_worker.py`: `WorkerProc._return_result()`, `_async_output_loop()`, `_generate_async_output_id()`
+- `vllm_omni/diffusion/executor/multiproc_executor.py`: `ResultPumpThread`, `collective_rpc()` two-path dispatch, `wait_output_ready()`, `_batch_split_map`
+- `vllm_omni/diffusion/diffusion_engine.py`: `step()` / `step_streaming()` / `add_req_and_wait_for_response()` async output waiting
+- `vllm_omni/diffusion/sched/request_scheduler.py`: `update_from_output()` — `async_output_id` → `FINISHED_COMPLETED`
+- `vllm_omni/diffusion/ipc.py`: `pack_diffusion_output_shm()` side-stream D2H path
+- `vllm_omni/platforms/interface.py`: `OmniPlatform.record_device_event()` base class
+- `vllm_omni/platforms/cuda/platform.py`: `CudaOmniPlatform.record_device_event()`
+- `vllm_omni/platforms/npu/platform.py`: `NPUOmniPlatform.record_device_event()`
+- `vllm_omni/diffusion/models/hunyuan_image3/pipeline_hunyuan_image3.py`: `get_hunyuan_image3_post_process_func()`, postprocess removal
+- `vllm_omni/diffusion/registry.py`: `_DIFFUSION_POST_PROCESS_FUNCS` postprocess registry

docs/design/feature/omni_async_output_materialization.md

@@ -0,0 +1,386 @@
+# Async Omni Output Materialization
+
+## Table of Contents
+
+1. [Overview](#overview)
+2. [Performance](#performance)
+3. [Architecture](#architecture)
+4. [Enabling the Feature](#enabling-the-feature)
+5. [Compatibility and Fallbacks](#compatibility-and-fallbacks)
+6. [Related Files](#related-files)
+
+## Overview
+
+Async Omni output materialization moves CPU-side construction of
+`OmniModelRunnerOutput` out of the autoregressive (AR) decode critical path.
+It complements [Async Chunk](async_chunk.md): async chunk pipelines partial
+outputs between stages, while async output materialization prevents the
+construction of those partial outputs from blocking the next decode step.
+
+An AR stage must return sampled token IDs quickly because the scheduler needs
+them for the next decode step. However, an Omni stage can also produce hidden
+states, multimodal tensors, streaming inter-stage payloads, and connector
+metadata. Building these payloads can require device-to-host (D2H) copies,
+tensor slicing, flattening, and Python object construction.
+
+Without this feature, all of that work runs inline in `sample_tokens()`:
+
+```text
+sample tokens
+  -> update decode state
+  -> copy hidden and multimodal outputs to CPU
+  -> build per-request payloads
+  -> build the streaming wire payload
+  -> collect connector signals
+  -> return sampled tokens
+  -> launch the next decode step
+```
+
+With async output materialization, only work required by the next decode step
+stays on the critical path:
+
+```text
+sample tokens
+  -> update decode state
+  -> snapshot output state and start asynchronous D2H copies
+  -> register sampled tokens
+  -> launch the next decode step
+
+background output path
+  -> wait for the payload snapshot
+  -> build per-request payloads
+  -> build the streaming wire payload
+  -> collect connector signals
+  -> construct OmniModelRunnerOutput
+```
+
+This allows payload construction for step `N` to overlap with GPU decode work
+for step `N + 1`. It does not change model computation or generated output.
+
+!!! note "Platform and validation scope"
+    The performance and correctness validation in this document covers CUDA
+    and ROCm. This is a validation scope, not a runtime platform guard: XPU
+    inherits `GPUARModelRunner`, and MUSA selects `GPUARWorker`, so those
+    backends may enter the asynchronous path when its other guards pass. XPU
+    and MUSA have not been validated for this optimization. Ascend NPU uses a
+    separate `NPUARModelRunner` path that still constructs
+    `OmniModelRunnerOutput` and drains connector output synchronously before
+    wrapping the result as an asynchronous runner output.
+
+<p align="center">
+  <img
+    alt="Decode step gap before and after async output materialization"
+    src="../../../source/architecture/qwen3-omni-async-output-step-gap.svg"
+    width="100%"
+  >
+</p>
+
+<p align="center">
+  <em>
+    Figure 1: Moving payload construction off the decode path reduces the
+    observed gap between consecutive Talker steps from about 2.8 ms to 41 µs.
+    Adapted from the
+    <a href="https://vllm.ai/blog/2026-07-01-qwen3-omni-optimization">
+      Qwen3-Omni optimization blog
+    </a>.
+  </em>
+</p>
+
+For Qwen3-Omni, the feature is enabled for both AR stages:
+
+- **Thinker**: Defers hidden-state and multimodal payload materialization.
+- **Talker**: Runs its decode-state postprocess eagerly, then defers codec
+  payload materialization. Because Code2Wav consumes codec codes rather than
+  Talker hidden states, the Talker also skips unnecessary hidden-state D2H.
+- **Code2Wav**: Does not use this path because it is a generation stage rather
+  than an AR stage.
+
+Qwen3-TTS uses the same mechanism in its AR Talker stage. The Talker updates
+the decode state it needs for the next step before returning, then constructs
+the codec payload for Code2Wav in the background.
+
+## Performance
+
+In the controlled Qwen3-Omni optimization sweep, enabling async output
+materialization on top of CUDA Graph and async chunk produced the following
+results at concurrency 64:
+
+| Configuration | Request throughput | Mean audio TTFP | Mean audio RTF |
+|---|---:|---:|---:|
+| CUDA Graph + async chunk | 9.3 req/s | 655 ms | 0.63 |
+| + async output materialization | 11.3 req/s | 631 ms | 0.47 |
+| Change | **+22%** | **-4%** | **-25%** |
+
+The sweep used `Qwen/Qwen3-Omni-30B-A3B-Instruct`, 640 Seed-TTS English
+prompts, and one replica each for Thinker, Talker, and Code2Wav. Async output
+materialization recovers throughput by removing CPU payload construction from
+the decode critical path without regressing time to first audio packet (TTFP).
+See the
+[Qwen3-Omni optimization blog](https://vllm.ai/blog/2026-07-01-qwen3-omni-optimization)
+for the complete benchmark methodology and results.
+
+The implementation was also validated against the synchronous safe-copy path.
+The text and audio output hashes matched, confirming that overlap does not
+change the generated result. See
+[PR #4476](https://github.com/vllm-project/vllm-omni/pull/4476) for the
+correctness and profiling details.
+
+## Architecture
+
+### Output Lifecycle
+
+`OmniAsyncGPUModelRunnerOutput` extends vLLM's
+`AsyncGPUModelRunnerOutput`. It preserves the existing asynchronous
+sampled-token feedback while adding a background builder for the complete Omni
+output.
+
+The output lifecycle is:
+
+1. `GPUARModelRunner.sample_tokens()` samples tokens and performs the
+   bookkeeping required by the next decode step.
+2. The runner snapshots step-local metadata such as request IDs, token spans,
+   scheduler output, and query start locations.
+3. CUDA payload tensors are cloned before reusable model or CUDA Graph buffers
+   can be overwritten.
+4. The cloned payload is copied to pinned CPU memory on a dedicated CUDA
+   stream. A CUDA event records when the copy is ready.
+5. `OmniAsyncGPUModelRunnerOutput` starts a background thread that waits for
+   the payload event and builds `OmniModelRunnerOutput`.
+6. The AR runner immediately registers the sampled-token CPU copy with the
+   input batch and returns, allowing async scheduling to advance.
+7. When the engine calls `get_output()`, it joins the background builder,
+   propagates any builder exception, and finalizes sampled tokens and logprobs
+   through the upstream output implementation.
+
+```text
+GPU / decode path                         Background output path
+
+forward + sample
+      |
+      +-- clone output tensors
+      +-- enqueue D2H copy --------------> wait for D2H event
+      +-- register sampled tokens               |
+      +-- return async output                    +-- slice per request
+      |                                         +-- build multimodal payloads
+next decode step                                +-- partition streaming payloads
+                                                +-- build the wire payload
+                                                +-- drain connector signals
+                                                +-- build OmniModelRunnerOutput
+```
+
+### Safe State Snapshots
+
+The next scheduler step can mutate runner state while the background builder is
+still active. Most step-local builder inputs are therefore detached snapshots.
+The snapshotted state includes:
+
+- scheduler token counts and speculative-token metadata
+- request IDs and request-to-batch-index mappings
+- sampled token IDs, logprobs, and prompt logprobs
+- query start locations and scheduled-token spans
+- hidden states and multimodal output tensors
+- KV-connector and encoder-cache outputs captured for the step
+
+CUDA tensor cloning is required because CUDA Graph and model output buffers can
+be reused by the next forward pass. The dedicated copy stream and pinned host
+buffers make the D2H transfer asynchronous; the snapshot retains its cloned
+CUDA sources until the transfer event completes.
+
+### Live Runner State and Ownership
+
+The complete builder input is not a detached snapshot. Connector output state
+remains runner-owned: `get_omni_connector_output()` drains live, per-cycle
+connector signals after the streaming payload is built. Correctness therefore
+depends on that builder being the only consumer that drains connector signals
+for its output cycle. Receive-side connector state written by the background
+receiver is coordinated by the connector mixin's `_lock`.
+`OmniAsyncGPUModelRunnerOutput.get_output()` joins the builder and is the
+completion and exception boundary before the materialized output is consumed.
+
+The shared output builder also contains a full-payload accumulation branch
+that reads `self.requests`. That branch is not reachable when async Omni output
+materialization is enabled: this feature requires `async_chunk`, while
+`should_accumulate_full_payload_output()` returns `False` whenever
+`async_chunk` is enabled. The feature path partitions each step's payload into
+streaming inter-stage and client handoffs instead.
+
+### Output Builder
+
+The background builder performs the work that previously ran inline:
+
+- resolves which requests need downstream payloads
+- converts hidden and multimodal tensors into per-request CPU payloads
+- applies model-specific payload processing
+- partitions per-step payloads into inter-stage and client streams
+- creates the tensor-only `multimodal_outputs` wire payload
+- drains live connector readiness signals after building the streaming payload
+- constructs the final `OmniModelRunnerOutput`
+
+Input-side connector operations remain synchronous with model execution. In
+particular, receiving inputs and flushing previously completed connector
+outputs are not deferred because they affect scheduler-visible request state.
+
+### Qwen3-Omni Stage Behavior
+
+| Stage | Async output behavior | Reason |
+|---|---|---|
+| Thinker | Snapshots hidden states and multimodal outputs; builds the downstream payload in the background | The Talker needs the payload, but the next Thinker decode step only needs sampled-token feedback |
+| Talker | Runs lightweight postprocess eagerly; snapshots codec outputs; omits hidden states from the downstream payload | `hidden_states.last` is needed by the next Talker step, while Code2Wav only needs codec codes |
+| Code2Wav | Uses the normal generation-stage output path | Code2Wav is not executed by `GPUARModelRunner` |
+
+The Qwen3-TTS Talker follows the same pattern as the Qwen3-Omni Talker:
+postprocess runs eagerly because the next decode step needs its updated state,
+while codec payload construction runs asynchronously. Its Code2Wav stage also
+uses the normal generation-stage path.
+
+## Enabling the Feature
+
+There is no separate `async_omni_output` command-line or YAML option. The
+feature is selected automatically for model stages that opt in when the runtime
+conditions are safe. On the validated CUDA and ROCm configurations, using the
+bundled deployment profile is enough to turn it on for the supported
+Qwen3-Omni and Qwen3-TTS pipelines. The selection logic has no CUDA/ROCm-only
+guard; see [Compatibility and Fallbacks](#compatibility-and-fallbacks) for the
+other platform paths. The same profile does not enable background Omni
+materialization on Ascend NPU.
+
+### Example 1: Qwen3-Omni
+
+Start the full Qwen3-Omni pipeline:
+
+```bash
+vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct \
+  --omni \
+  --port 8091
+```
+
+The model registry automatically loads
+`vllm_omni/deploy/qwen3_omni_moe.yaml`. Its relevant settings are:
+
+```yaml
+async_chunk: true
+
+stages:
+  - stage_id: 0  # Thinker
+    enable_prefix_caching: false
+    # async_scheduling defaults to true for AR stages
+
+  - stage_id: 1  # Talker
+    enable_prefix_caching: false
+    # async_scheduling defaults to true for AR stages
+
+  - stage_id: 2  # Code2Wav
+    enable_prefix_caching: false
+    async_scheduling: false
+```
+
+With this profile, async output materialization activates automatically for:
+
+- **Stage 0, Thinker**: Defers hidden-state and multimodal payload
+  construction.
+- **Stage 1, Talker**: Defers codec payload construction after eagerly
+  updating its decode state.
+
+Stage 2 does not use this feature because Code2Wav is not an AR stage.
+
+### Example 2: Qwen3-TTS
+
+Start Qwen3-TTS with any checkpoint that uses the Qwen3-TTS Talker pipeline.
+For example:
+
+```bash
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+  --omni \
+  --port 8091
+```
+
+The model registry automatically loads `vllm_omni/deploy/qwen3_tts.yaml`. Its
+relevant settings are:
+
+```yaml
+async_chunk: true
+
+stages:
+  - stage_id: 0  # Talker
+    async_scheduling: true
+    enable_prefix_caching: false
+
+  - stage_id: 1  # Code2Wav
+    enable_prefix_caching: false
+```
+
+Async output materialization activates only for Stage 0, the AR Talker. The
+Talker keeps the state required by its next decode step on the GPU, skips an
+unnecessary hidden-state D2H copy, and builds the codec payload for Code2Wav in
+the background. Stage 1 uses the generation-stage output path.
+
+The same behavior applies to the Base and VoiceDesign Qwen3-TTS checkpoints
+because they use the same Talker implementation.
+
+!!! warning
+    Do not pass `--no-async-chunk` or enable prefix caching when you want this
+    optimization. Either change causes the runner to fall back to synchronous
+    output construction. Enabling async chunk alone cannot activate the feature
+    for a model that has not opted into async Omni output.
+
+!!! note "Prefix cache compatibility"
+    Async Omni output materialization and Omni prefix caching cannot currently
+    run together. `_should_use_async_omni_output()` returns `False` whenever
+    `self.omni_prefix_cache` is present, so prefix caching selects synchronous
+    output materialization. Supporting both features requires snapshotting or
+    otherwise synchronizing prefix-cache merge and update state before the
+    background output path can consume it safely.
+
+## Compatibility and Fallbacks
+
+`GPUARModelRunner` uses the async output path only when all of the following
+runtime conditions hold:
+
+| Requirement | Reason |
+|---|---|
+| AR async scheduling is enabled | The optimization relies on the scheduler advancing while the prior output is materialized |
+| `async_chunk` is enabled | The feature targets incremental downstream Omni payloads |
+| The model stage opts in with `use_async_omni_output` | Models must declare that their output lifecycle is safe to defer |
+| Omni prefix cache is disabled | Prefix-cache merge and update ordering currently requires synchronous materialization |
+| Speculative decoding is disabled | Speculative output state is not included in this deferred path |
+| Routed-expert output is disabled | Routed-expert extraction currently requires the synchronous path |
+| Postprocess is absent or explicitly runs eagerly | State needed by the next decode step must be updated before the runner returns |
+
+These checks are evaluated per stage on runners based on `GPUARModelRunner`.
+An unsupported combination does not prevent serving; it only falls back to
+synchronous output construction for that stage. CUDA and ROCm are the validated
+platforms. XPU inherits `GPUARModelRunner`, and MUSA selects `GPUARWorker`, so
+they may select the async path when these checks pass, but this optimization is
+not yet validated on either backend.
+
+Ascend NPU does not use this selection path. `NPUARModelRunner.sample_tokens()`
+fully constructs `OmniModelRunnerOutput`, calls
+`get_omni_connector_output()`, and only then creates the asynchronous wrapper,
+so Omni payload materialization remains synchronous.
+
+!!! note
+    `use_async_omni_output`,
+    `eager_omni_postprocess_before_async_output`, and
+    `omni_pooler_payload_include_hidden` are model implementation contracts,
+    not user-facing configuration fields.
+
+## Related Files
+
+- `vllm_omni/worker/gpu_ar_model_runner.py`: Async output object, safe tensor
+  snapshots, compatibility guards, and deferred Omni output builder.
+- `vllm_omni/platforms/npu/worker/npu_ar_model_runner.py`: Separate Ascend NPU
+  runner, where Omni output materialization remains synchronous.
+- `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py`: Thinker and
+  Talker opt-in behavior.
+- `vllm_omni/deploy/qwen3_omni_moe.yaml`: Default Qwen3-Omni deployment
+  settings.
+- `vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py`: Qwen3-TTS
+  Talker opt-in and eager postprocess behavior.
+- `vllm_omni/deploy/qwen3_tts.yaml`: Default Qwen3-TTS deployment settings.
+- `tests/worker/test_gpu_ar_model_runner.py`: Snapshot, guard, connector
+  ordering, and background error-propagation tests.
+- [Async Chunk](async_chunk.md): Inter-stage chunking and scheduling design.
+- [Qwen3-Omni optimization blog](https://vllm.ai/blog/2026-07-01-qwen3-omni-optimization):
+  Optimization context and controlled performance results.
+- [PR #4476](https://github.com/vllm-project/vllm-omni/pull/4476): Feature
+  implementation, validation, and review history.

docs/design/feature/tensor_parallel.md

@@ -264,7 +264,7 @@ Complete examples in the codebase:
 | **Z-Image** | `vllm_omni/diffusion/models/z_image/z_image_transformer.py` | Standard TP | Full implementation with validation |
 | **FLUX** | `vllm_omni/diffusion/models/flux/flux_transformer.py` | Dual-stream | Image + text streams |
 | **Qwen-Image** | `vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py` | Standard TP | With RoPE |
-| **TP Tests** | `tests/e2e/offline_inference/test_zimage_parallelism.py` | E2E testing | TP correctness and performance |
+| **TP Tests** | `tests/diffusion/distributed/test_tensor_parallel.py` | E2E testing | TP correctness and performance |
 | **Constraint Tests** | `tests/diffusion/models/z_image/test_zimage_tp_constraints.py` | Unit testing | Validation logic |
 
 ---

docs/design/index.md

@@ -10,6 +10,7 @@ This section contains design documents and architecture specifications for vLLM-
 
 - [Disaggregated Inference](feature/disaggregated_inference.md)
 - [Ray-based Execution](feature/ray_based_execution.md)
+- [Async Omni Output Materialization](feature/omni_async_output_materialization.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)

docs/design/metrics.md

@@ -148,7 +148,13 @@ Upstream vLLM's `Scheduler.make_stats()` runs on every AR generation step, retur
 
 `OmniSchedulerMixin.make_stats()` (in `vllm_omni/core/sched/omni_scheduler_mixin.py`) throttles stats emission to at most once per second **per scheduler** — i.e. per `(stage, replica)` since each replica owns its own scheduler instance. Between intervals it returns `None`, which the engine core skips serializing. This keeps gauges fresh enough for Prometheus scrapes (typically 15-30s intervals) while eliminating the per-step overhead.
 
-The orchestrator side does not add its own throttle on top: the per-replica recording loop gates only on `raw_outputs.scheduler_stats is not None` (i.e. this replica's scheduler passed its own 1 Hz gate). A previous global `_last_stats_ts` on the orchestrator was removed because it starved every replica other than the first to emit within each second.
+The orchestrator side does not add its own throttle on top. Scheduler gauges are
+recorded when `raw_outputs.scheduler_stats is not None` (i.e. this replica's
+scheduler passed its own 1 Hz gate), while output-batch request/token metrics
+are recorded from `IterationStats` whenever a non-empty output batch is
+processed. A previous global `_last_stats_ts` on the orchestrator was removed
+because it starved every replica other than the first to emit within each
+second.
 
 ## Metric Definitions
 

docs/design/qwen3_omni_tts_performance_optimization.md

@@ -419,8 +419,7 @@ Notes:
 ```bash
 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct \
   --omni \
-  --port 8091 \
-  --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_omni_moe_async_chunk.yaml
+  --port 8091
 ```
 
 #### 3) Key config knobs
@@ -494,7 +493,7 @@ The default config (`qwen3_tts.yaml`) enables the full optimization stack:
 vllm-omni serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
   --omni \
   --port 8000 \
-  --stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts_no_async_chunk.yaml
+  --no-async-chunk
 ```
 
 #### 3) Key config knobs

docs/getting_started/installation/README.md

@@ -1,5 +1,8 @@
 # Installation
 
+!!! important
+    vLLM-Omni is released against the matching upstream vLLM major/minor version. vLLM-Omni 0.26.x requires vLLM 0.26.x; the stable v0.26.0 instructions pin vLLM 0.26.0.
+
 vLLM-Omni supports the following hardware platforms:
 
 - [GPU](gpu.md)

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

@@ -5,7 +5,7 @@
 # --8<-- [end:requirements]
 # --8<-- [start:set-up-using-python]
 
-vLLM-Omni depends vLLM. So please follow instructions below mainly for vLLM.
+vLLM-Omni depends on the matching major/minor release of vLLM. The vLLM-Omni 0.26.x release line uses vLLM 0.26.x.
 
 !!! note
     PyTorch installed via `conda` will statically link `NCCL` library, which can cause issues when vLLM tries to use `NCCL`. See <gh-issue:8420> for more details.
@@ -124,7 +124,7 @@ If you want to specify the base vLLM version:
 ```bash
 DOCKER_BUILDKIT=1 docker build \
   -f docker/Dockerfile.cuda \
-  --build-arg BASE_IMAGE=vllm/vllm-openai:v0.22.1 \
+  --build-arg BASE_IMAGE=vllm/vllm-openai:v0.26.0 \
   -t vllm-omni-cuda .
 ```
 

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

@@ -9,8 +9,8 @@ The recommended way to use vLLM-Omni on NPU is through the vLLM-Ascend pre-built
 
 ```bash
 # vLLM-Ascend image
-# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.25.0
-# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.25.0-a3
+# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.26.0
+# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.26.0-a3
 docker run --rm \
     --name vllm-omni-npu \
     --shm-size=1g \
@@ -28,11 +28,11 @@ docker run --rm \
     -v /etc/ascend_install.info:/etc/ascend_install.info \
     -v /root/.cache:/root/.cache \
     -p 8000:8000 \
-    -it quay.io/atlas-ci/vllm-ascend:0.25.0-a3 bash
+    -it quay.io/atlas-ci/vllm-ascend:v0.26.0-a3 bash
 
 # Inside the container, install vLLM-Omni from source
 cd /vllm-workspace
-git clone -b v0.25.0 https://github.com/vllm-project/vllm-omni.git
+git clone -b v0.26.0 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 .
@@ -53,16 +53,20 @@ We are keeping [issue #886](https://github.com/vllm-project/vllm-omni/issues/886
 You can also build vLLM-Omni from the latest main branch if you want to use the latest features or bug fixes. (But sometimes it will break for a while. You can check [issue #886](https://github.com/vllm-project/vllm-omni/issues/886) for the status of the latest commit of vLLM-Omni main branch on NPU.)
 
 ```bash
-# Pin vLLM version to 0.18.0
-git clone -b v0.18.0 https://github.com/vllm-project/vllm.git
+# Pin vLLM and vLLM-Ascend to the v0.26 release line
+git clone -b v0.26.0 https://github.com/vllm-project/vllm.git
+cd vllm
 VLLM_TARGET_DEVICE=empty pip install -v -e .
+cd ..
 
-git clone -b v0.18.0rc1 https://github.com/vllm-project/vllm-ascend.git
+git clone -b releases/v0.26.0rc https://github.com/vllm-project/vllm-ascend.git
+cd vllm-ascend
 pip install -v -e .
+cd ..
 
 # Install vLLM-Omni from the latest main branch
 git clone https://github.com/vllm-project/vllm-omni.git
-cd /vllm-workspace/vllm-omni
+cd vllm-omni
 pip install -v -e . --no-build-isolation
 # or VLLM_OMNI_TARGET_DEVICE=npu pip install -v -e .
 export VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -84,8 +88,8 @@ Supported images as following.
 Here's an example deployment command that has been verified on 4 x NPUs:
 
 ```bash
-# Atlas A2: quay.io/ascend/vllm-omni:v0.25.0
-# Atlas A3: quay.io/ascend/vllm-omni:v0.25.0-a3
+# Atlas A2: quay.io/ascend/vllm-omni:v0.26.0
+# Atlas A3: quay.io/ascend/vllm-omni:v0.26.0-a3
 docker run --rm \
     --name vllm-omni-a3 \
     --shm-size=4g \
@@ -110,7 +114,7 @@ docker run --rm \
     -v /etc/ascend_install.info:/etc/ascend_install.info \
     -v ~/.cache:/root/.cache \
     -p 8091:8091 \
-    -it quay.io/ascend/vllm-omni:v0.25.0-a3 bash
+    -it quay.io/ascend/vllm-omni:v0.26.0-a3 bash
 ```
 
 !!! tip
@@ -122,9 +126,9 @@ Or build IMAGE from **source code**:
 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
 # A2
-docker build -t vllm-omni-dev-image:latest -f ./docker/Dockerfile.npu .
+# docker build -t vllm-omni-dev-image:latest -f ./docker/Dockerfile.npu .
 # A3
-# docker build -t vllm-omni-dev-image:latest -f ./docker/Dockerfile.npu.a3 .
+docker build -t vllm-omni-dev-image:latest -f ./docker/Dockerfile.npu.a3 .
 ```
 
 # --8<-- [end:pre-built-images]

docs/models/supported_models.md

@@ -44,6 +44,7 @@ th {
 | `LTX2Pipeline` | LTX-2 / LTX-2.3 one-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
 | `LTX2DistilledPipeline` | LTX-2 distilled two-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | |
 | `LingBotVideoPipeline` | LingBot-Video dense and MoE T2V | `robbyant/lingbot-video-dense-1.3b`, `robbyant/lingbot-video-moe-30b-a3b` | ✅︎ | | | |
+| `MiniMaxH3Pipeline` | MiniMax H3 T2VA, FL2VA, and Ref2VA | `MiniMaxAI/MiniMax-H3` | ✅︎ | | | |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | |
 | `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | |

docs/serving/speech_api.md

@@ -276,20 +276,42 @@ Client -> Server:
 
 | Message | Description |
 |---------|-------------|
-| `{"type": "session.config", ...}` | Session configuration (sent once, first message) |
+| `{"type": "session.config", ...}` | Session configuration (first message; may be resent between utterances to change it) |
 | `{"type": "input.text", "text": "..."}` | Text chunk |
-| `{"type": "input.done"}` | End of input, flushes remaining buffer |
+| `{"type": "input.done"}` | End of utterance: flushes the buffer and keeps the connection open |
+| `{"type": "session.close"}` | End of connection |
 
 Server -> Client:
 
 | Message | Description |
 |---------|-------------|
-| `{"type": "audio.start", "sentence_index": 0, "sentence_text": "...", "format": "pcm", "sample_rate": 24000}` | Audio generation starting for the buffered input |
+| `{"type": "audio.start", "utterance_index": 0, "sentence_index": 0, "sentence_text": "...", "format": "pcm", "sample_rate": 24000}` | Audio generation starting for the buffered input |
 | Binary frame | Raw audio bytes (one or more PCM chunks when `stream_audio=true`) |
-| `{"type": "audio.done", "sentence_index": 0, "total_bytes": 96000, "error": false}` | Audio complete for the buffered input |
-| `{"type": "session.done", "total_sentences": N}` | Session complete |
+| `{"type": "audio.done", "utterance_index": 0, "sentence_index": 0, "total_bytes": 96000, "error": false}` | Audio complete for the buffered input |
+| `{"type": "session.done", "utterance_index": 0, "total_sentences": N}` | Flushed utterance complete |
 | `{"type": "error", "message": "..."}` | Non-fatal error |
 
+### Flushing vs. Closing
+
+`input.done` is a flush, not a disconnect. The server synthesizes the buffered
+text, emits `session.done`, and then waits on the same connection for the next
+utterance, so a client that speaks repeatedly (for example one driven by an
+upstream LLM) pays the WebSocket handshake once instead of once per utterance.
+
+* The session config is sticky. Send `input.text` again straight after
+  `session.done` to reuse it, or send another `session.config` first to change
+  voice, format, or reference audio. A `session.config` sent while text is
+  still buffered is rejected so no pending input is silently dropped.
+* An utterance is the flush unit, not a linguistic one: it is whatever text was
+  buffered when `input.done` arrived, of any length, synthesized as one request.
+  `utterance_index` counts those flushes across the connection, so it tells you
+  which `input.done` a frame belongs to. `sentence_index` counts within one
+  flush and so pairs with `total_sentences`, which means every utterance reports
+  `sentence_index: 0` of `total_sentences: 1` (or `0` for an empty buffer).
+* End the connection with `session.close`, or by closing the socket. An idle
+  connection is still closed after the server's idle timeout, which now also
+  applies to the gap between utterances.
+
 ### Session Config Parameters
 
 All REST API parameters are supported, plus:

docs/source/architecture/qwen3-omni-async-output-step-gap.svg

@@ -0,0 +1,150 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="1024" height="600" viewBox="0 0 1024 600" role="img" aria-labelledby="title desc">
+  <title id="title">Qwen3-Omni Talker step gap before and after asynchronous output materialization</title>
+  <desc id="desc">Before the optimization, synchronous D2H copy, per-request payload construction, streaming payload handoff, and connector-signal drain separate consecutive Talker forward and sample steps by about 2.8 milliseconds. After the optimization, that Omni output work runs on a background builder while the next Talker step executes, leaving a boundary of about 41 microseconds for snapshot, D2H enqueue, state update, and sampled-token feedback.</desc>
+  <defs>
+    <marker id="time-arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto">
+      <path d="M 0 0 L 10 5 L 0 10 z" fill="#64748b"/>
+    </marker>
+    <marker id="red-end" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="5" markerHeight="5" orient="auto">
+      <path d="M 0 0 L 10 5 L 0 10 z" fill="#b91c1c"/>
+    </marker>
+    <marker id="red-start" viewBox="0 0 10 10" refX="2" refY="5" markerWidth="5" markerHeight="5" orient="auto">
+      <path d="M 10 0 L 0 5 L 10 10 z" fill="#b91c1c"/>
+    </marker>
+    <marker id="blue-end" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="6" markerHeight="6" orient="auto">
+      <path d="M 0 0 L 10 5 L 0 10 z" fill="#2563eb"/>
+    </marker>
+    <style>
+      .canvas { fill: #ffffff; stroke: #e5e7eb; }
+      .title { font: 700 22px system-ui, -apple-system, Segoe UI, sans-serif; fill: #111827; }
+      .subtitle { font: 500 13px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
+      .phase { font: 700 16px system-ui, -apple-system, Segoe UI, sans-serif; fill: #1f2937; }
+      .phase-sub { font: 500 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
+      .lane { font: 600 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #374151; }
+      .block-title { font: 700 10px system-ui, -apple-system, Segoe UI, sans-serif; letter-spacing: .25px; fill: #ffffff; }
+      .block-op { font: 500 9.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
+      .axis { font: 500 11px system-ui, -apple-system, Segoe UI, sans-serif; fill: #94a3b8; }
+      .measure { font: 700 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #b91c1c; }
+      .inset-title { font: 700 11px system-ui, -apple-system, Segoe UI, sans-serif; fill: #475569; }
+      .inset-step { font: 600 10px system-ui, -apple-system, Segoe UI, sans-serif; fill: #ffffff; }
+      .note-title { font: 700 12px system-ui, -apple-system, Segoe UI, sans-serif; fill: #1e3a8a; }
+      .note-text { font: 500 10.5px system-ui, -apple-system, Segoe UI, sans-serif; fill: #475569; }
+      .stat-cap { font: 600 11px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
+      .stat-big { font: 800 16px system-ui, -apple-system, Segoe UI, sans-serif; fill: #111827; }
+      .stat-sub { font: 500 11px system-ui, -apple-system, Segoe UI, sans-serif; fill: #64748b; }
+      .before-card { fill: #eef4ff; stroke: #c7dbf6; }
+      .after-card { fill: #eafaf0; stroke: #bfe7cf; }
+      .talker { fill: #16a34a; }
+      .sync-output { fill: #ea580c; }
+      .async-output { fill: #2563eb; }
+      .zoom-card { fill: #fff7f7; stroke: #f3caca; }
+      .note-card { fill: #eff6ff; stroke: #bfdbfe; }
+      .stat-blue { fill: #eef4ff; stroke: #c7dbf6; }
+      .stat-green { fill: #eafaf0; stroke: #bfe7cf; }
+      .stat-gray { fill: #f4f5f7; stroke: #e2e5ea; }
+      .guide { stroke: #e2e8f0; stroke-width: 1; }
+      .time-line { stroke: #64748b; stroke-width: 1.3; fill: none; marker-end: url(#time-arrow); }
+      .measure-line { stroke: #b91c1c; stroke-width: 1.4; marker-start: url(#red-start); marker-end: url(#red-end); }
+      .zoom-lead { stroke: #dc2626; stroke-width: 1.15; stroke-dasharray: 4 3; fill: none; }
+      .handoff { stroke: #2563eb; stroke-width: 1.3; stroke-dasharray: 3 2; fill: none; marker-end: url(#blue-end); }
+    </style>
+  </defs>
+
+  <rect x="14" y="14" width="996" height="572" rx="14" class="canvas"/>
+
+  <text x="44" y="52" class="title">Async output: what moves off the Talker decode path</text>
+  <text x="44" y="74" class="subtitle">Qwen3-Omni Talker stage &#183; consecutive decode steps in a fixed profiler window</text>
+
+  <!-- ===================== BEFORE ===================== -->
+  <rect x="44" y="94" width="144" height="110" rx="10" class="before-card"/>
+  <text x="60" y="120" class="phase">Before</text>
+  <text x="60" y="139" class="phase-sub">Omni output built</text>
+  <text x="60" y="156" class="phase-sub">synchronously on the</text>
+  <text x="60" y="173" class="phase-sub">AR critical path</text>
+
+  <text x="206" y="122" class="lane">AR path</text>
+  <line x1="300" y1="88" x2="300" y2="210" class="guide"/>
+
+  <rect x="300" y="104" width="190" height="54" rx="5" class="talker"/>
+  <text x="395" y="124" text-anchor="middle" class="block-title">TALKER STAGE &#183; STEP N</text>
+  <text x="395" y="145" text-anchor="middle" class="block-op">forward &#8594; sample</text>
+
+  <rect x="490" y="104" width="250" height="54" rx="5" class="sync-output"/>
+  <text x="615" y="120" text-anchor="middle" class="block-title">SYNC OMNI OUTPUT &#183; STEP N</text>
+  <text x="615" y="138" text-anchor="middle" class="block-op">D2H &#8594; per-request slice / payload build</text>
+  <text x="615" y="152" text-anchor="middle" class="block-op">streaming handoff &#8594; connector-signal drain</text>
+
+  <rect x="740" y="104" width="220" height="54" rx="5" class="talker"/>
+  <text x="850" y="124" text-anchor="middle" class="block-title">TALKER STAGE &#183; STEP N+1</text>
+  <text x="850" y="145" text-anchor="middle" class="block-op">forward &#8594; sample</text>
+
+  <line x1="490" y1="174" x2="740" y2="174" class="measure-line"/>
+  <text x="615" y="194" text-anchor="middle" class="measure">~2.8 ms GPU gap: output work is inline</text>
+
+  <!-- ===================== AFTER ===================== -->
+  <rect x="44" y="218" width="144" height="140" rx="10" class="after-card"/>
+  <text x="60" y="244" class="phase">After</text>
+  <text x="60" y="263" class="phase-sub">Only step feedback</text>
+  <text x="60" y="280" class="phase-sub">stays inline; payload</text>
+  <text x="60" y="297" class="phase-sub">work runs in the</text>
+  <text x="60" y="314" class="phase-sub">background</text>
+
+  <text x="206" y="254" class="lane">Talker GPU</text>
+  <text x="206" y="326" class="lane">Output builder</text>
+  <line x1="300" y1="212" x2="300" y2="370" class="guide"/>
+
+  <rect x="300" y="226" width="235" height="54" rx="5" class="talker"/>
+  <text x="418" y="246" text-anchor="middle" class="block-title">TALKER STAGE &#183; STEP N</text>
+  <text x="418" y="267" text-anchor="middle" class="block-op">forward &#8594; sample</text>
+
+  <rect x="540" y="226" width="235" height="54" rx="5" class="talker"/>
+  <text x="658" y="246" text-anchor="middle" class="block-title">TALKER STAGE &#183; STEP N+1</text>
+  <text x="658" y="267" text-anchor="middle" class="block-op">forward &#8594; sample</text>
+
+  <path d="M 535 280 L 548 298" class="handoff"/>
+  <rect x="540" y="298" width="360" height="58" rx="5" class="async-output"/>
+  <text x="720" y="315" text-anchor="middle" class="block-title">BACKGROUND OMNI OUTPUT &#183; STEP N</text>
+  <text x="720" y="334" text-anchor="middle" class="block-op">wait for D2H &#8594; per-request slice / payload build</text>
+  <text x="720" y="349" text-anchor="middle" class="block-op">streaming handoff &#8594; connector-signal drain</text>
+
+  <line x1="300" y1="366" x2="930" y2="366" class="time-line"/>
+  <text x="922" y="382" text-anchor="end" class="axis">time</text>
+
+  <!-- leads from the real 41 us boundary to the magnified inset -->
+  <line x1="535" y1="280" x2="485" y2="424" class="zoom-lead"/>
+  <line x1="540" y1="280" x2="525" y2="424" class="zoom-lead"/>
+
+  <!-- ===================== 41 µs BOUNDARY ===================== -->
+  <rect x="300" y="386" width="400" height="92" rx="10" class="zoom-card"/>
+  <text x="316" y="404" class="inset-title">Zoom &#183; Talker step boundary (magnified)</text>
+
+  <rect x="320" y="416" width="165" height="26" rx="4" class="talker"/>
+  <rect x="525" y="416" width="155" height="26" rx="4" class="talker"/>
+  <text x="403" y="433" text-anchor="middle" class="inset-step">Talker step N</text>
+  <text x="603" y="433" text-anchor="middle" class="inset-step">Talker step N+1</text>
+
+  <line x1="485" y1="452" x2="525" y2="452" class="measure-line"/>
+  <text x="505" y="469" text-anchor="middle" class="measure">~41 &#181;s</text>
+
+  <rect x="720" y="386" width="240" height="92" rx="10" class="note-card"/>
+  <text x="736" y="405" class="note-title">Inline at the ~41 &#181;s boundary</text>
+  <text x="736" y="424" class="note-text">&#8226; eager Talker state update</text>
+  <text x="736" y="441" class="note-text">&#8226; output snapshot + D2H enqueue</text>
+  <text x="736" y="458" class="note-text">&#8226; sampled-token feedback</text>
+
+  <!-- ===================== SUMMARY ===================== -->
+  <rect x="44" y="494" width="296" height="70" rx="9" class="stat-blue"/>
+  <text x="60" y="514" class="stat-cap">Step-to-step gap</text>
+  <text x="60" y="538" class="stat-big">2.8 ms &#8594; 41 &#181;s</text>
+  <text x="60" y="555" class="stat-sub">~68&#215; tighter</text>
+
+  <rect x="356" y="494" width="296" height="70" rx="9" class="stat-green"/>
+  <text x="372" y="514" class="stat-cap">Talker GPU</text>
+  <text x="372" y="538" class="stat-big">next step starts immediately</text>
+  <text x="372" y="555" class="stat-sub">decode workers stay busy</text>
+
+  <rect x="668" y="494" width="312" height="70" rx="9" class="stat-gray"/>
+  <text x="684" y="514" class="stat-cap">Background builder</text>
+  <text x="684" y="538" class="stat-big">materializes Omni output</text>
+  <text x="684" y="555" class="stat-sub">payload work overlaps the next Talker step</text>
+</svg>

docs/user_guide/diffusion/attention_backends.md

@@ -15,7 +15,7 @@ The full set of backends and their platform defaults is in the **Backend Options
 | Value | Notes |
 |---|---|
 | `TRTLLM_ATTN` | FlashInfer's trtllm-gen FMHA (TensorRT-LLM's generated kernels, vendored by FlashInfer). BF16, GQA native, `head_dim=128`. Datacenter Blackwell only (sm_100 / sm_103). Supports **Skip-Softmax** sparse attention — see below. Requires `flashinfer`. |
-| `FLASH_ATTN` | Wraps FlashAttention 2. Default on Hopper / Ada / Ampere when `flash-attn` is installed. |
+| `FLASH_ATTN` | Wraps FlashAttention 4 on Blackwell when `flash-attn-4` is installed, then falls back to FlashAttention 3/2. Default on Hopper / Ada / Ampere when a compatible FlashAttention package is installed. |
 | `CUDNN_ATTN` | Pins `sdpa_kernel([CUDNN_ATTENTION])`. Default on Blackwell (sm_10x / sm_12x) with cuDNN ≥ 9.5. Wins on mask-heavy DiTs (HunyuanVideo-1.5: 2× e2e vs SDPA). |
 | `FLASHINFER_ATTN` | Calls FlashInfer's dense `single_prefill_with_kv_cache` directly with `custom_mask` for non-causal masked attention. Used as Blackwell fallback when cuDNN is unavailable. Requires `flashinfer`. |
 | `TORCH_SDPA` | PyTorch `scaled_dot_product_attention` with the default backend dispatcher. Most conservative; always available. |
@@ -167,6 +167,48 @@ sequence length, since Skip-Softmax only accelerates attention. Dense BF16 is `T
 Requires datacenter Blackwell with `head_dim == 128`; elsewhere, or without FlashInfer, selecting
 `TRTLLM_ATTN` raises rather than silently degrading.
 
+## TRTLLM_ATTN SAGE Quantization
+
+On top of dense/Skip-Softmax, `TRTLLM_ATTN` can run **SAGE** (SageAttention) — the trtllm-gen
+FMHA kernel with per-block quantized Q/K and per-channel fp8 V, so the two attention matmuls run
+in low precision. Enable it through the typed `quant` block on the attention spec:
+
+| Key | Valid values | Meaning |
+|---|---|---|
+| `dtype_qk` | `int8`, `fp8_e4m3` | Quantization dtype for Q and K. Absent ⇒ dense (SAGE off). |
+| `q_block_size` | `1`, `4`, `16` | Per-token block size for Q scales (default `1`). |
+| `k_block_size` | `1`, `4`, `16` | Per-token block size for K scales (default `16`). |
+
+V is always quantized per-channel to fp8_e4m3 and K-smoothing is applied inside the routine, so
+neither is a user knob. Only block sizes `1`, `4`, `16` have compiled kernels.
+
+The `quant` block is a shared `AttnQuantSpec` consumed by more than one backend; each reads the
+fields that apply and rejects the rest. `TRTLLM_ATTN` (SAGE) reads `dtype_qk` (`int8` / `fp8_e4m3`),
+`q_block_size`, and `k_block_size`; `FLASHINFER_ATTN` reads `dtype_qk` (`float16` / `bfloat16`),
+`dtype_vo`, and `flashinfer_backend` (see its section). A `dtype_qk` from the wrong set raises.
+
+```bash
+vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --diffusion-attention-config '{"default": {"backend": "TRTLLM_ATTN",
+      "quant": {"dtype_qk": "fp8_e4m3", "q_block_size": 1, "k_block_size": 16}}}'
+```
+
+```python
+from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, AttnQuantSpec
+
+AttentionConfig(
+    default=AttentionSpec(
+        backend="TRTLLM_ATTN",
+        quant=AttnQuantSpec(dtype_qk="fp8_e4m3", q_block_size=1, k_block_size=16),
+    ),
+)
+```
+
+
+**Requirements.** Needs FlashInfer ≥ 0.6.16rc1. Kernel
+availability is arch-dependent: `fp8_e4m3` QK has kernels on both **SM100** (B200) and **SM103**
+(B300); `int8` QK kernels are compiled for **SM100 only**.
+
 ## End-to-End Benchmark (BF16, sm_120 RTX Pro 6000 Blackwell)
 
 Same prompt and seed across runs. `Total generation time` from `text_to_video.py` / `text_to_image.py`.
@@ -193,9 +235,17 @@ DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN python examples/offline_inference/te
     --model Lightricks/LTX-2 ...
 ```
 
-### FA4 not yet integrated
+### FlashAttention-4 on Blackwell
 
-FlashAttention-4 (released March 2026) targets Blackwell natively and reportedly beats cuDNN by ~20% on B200. As of this writing the `flash-attn-4 4.0.0b10` wheel crashes with `AttributeError: 'NoneType' object has no attribute '_trait'` during JIT on sm_120. Not yet wired into vLLM-Omni; revisit when stable lands.
+Install the optional CUDA 13 extra to use the CuTe-based FA4 path:
+
+```bash
+pip install 'vllm-omni[fa4]'
+```
+
+`FLASH_ATTN` prefers `flash_attn.cute` on Blackwell and falls back to FA3/FA2
+when it is unavailable. Version `4.0.0b18` is required; earlier beta
+wheels had known JIT failures on Blackwell.
 
 ## Choosing a Backend Manually
 
@@ -353,64 +403,3 @@ DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3 python examples/offline_inference/text_t
     --tensor-parallel-size 2 \
     --output outputs/hv15_sage3.mp4
 ```
-
-### Mixed backends across roles
-
-Use `FLASH_ATTN` for self-attention and `TORCH_SDPA` for cross-attention:
-
-```bash
-python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
-    --prompt "A dog running across a field of golden wheat." \
-    --diffusion-attention-config.per_role.self.backend FLASH_ATTN \
-    --diffusion-attention-config.per_role.cross.backend TORCH_SDPA \
-    --tensor-parallel-size 2 \
-    --output outputs/wan22_mixed.mp4
-```
-
-### Compare against FlashAttention
-
-Unset the backend override, or explicitly use `FLASH_ATTN`:
-
-```bash
-python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 704 --width 1280 --num-frames 49 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 5.0 \
-    --tensor-parallel-size 2 \
-    --output outputs/wan22_fa3.mp4
-```
-
-## Validation Guidance
-
-Don't assume a faster attention backend is numerically interchangeable with `TORCH_SDPA`.
-
-Always compare:
-
-- End-to-end runtime
-- Diffusion-stage runtime (`add_req_and_wait` line in DiffusionEngine.step breakdown)
-- Output quality against a known-good baseline (CLIP similarity, frame-level diff, or visual review)
-
-At minimum, keep the same:
-
-- model
-- prompt
-- seed
-- resolution
-- frame count / step count
-- parallel config (TP / CFG-parallel / Ulysses degrees)
-
-## Reproducing the Benchmark Table
-
-The end-to-end numbers above were collected by running `text_to_video.py` /
-`text_to_image.py` with the same prompt and seed while varying
-`DIFFUSION_ATTENTION_BACKEND`. For a quick kernel-level comparison of the
-backends without loading a model:
-
-```bash
-python benchmarks/diffusion/bench_attention_backends.py --preset hv15
-```
-
-It runs all three BF16 backends on representative DiT attention shapes and
-prints a ranking table at the end.

docs/user_guide/diffusion/cpu_offload_diffusion.md

@@ -179,7 +179,134 @@ class Flux2Transformer2DModel(nn.Module):
 - Support single GPU only for now
 
 
-### Implementation Notes
+## Distributed Layerwise Offloading
+
+### How It Works
+
+Distributed layerwise offloading extends single-GPU layerwise offloading to
+multi-device deployments.  Each DP rank stores only **1/dp_size** of the model
+weights in host memory; full layer weights are reconstructed at runtime via
+**AllGather** on a dedicated communication stream, overlapped with computation
+via a fixed double-buffer scheme.
+
+**Key features:**
+- **Weight sharding + AllGather**: each rank stores 1/dp_size of weights,
+  reconstructed per-layer via `all_gather_into_tensor`
+- **Fixed double-buffer**: exactly 2 transformer blocks on each device at any
+  time, regardless of model size
+- **DP multi-concurrency**: N concurrent requests processed in parallel
+  (AllGather only gathers weight shards, which are request-independent)
+- **mmap weight loading**: weights loaded as mmap views pointing to shared OS
+  page cache, eliminating O(dp_size × model_size) RSS during model creation
+- **Platform-agnostic**: works on NVIDIA GPU (CUDA/NCCL) and Ascend NPU
+  (CANN/HCCL) via vLLM-Omni's platform abstraction
+
+**Execution Flow (per device):**
+
+```
+Compute Stream:  [Layer N]          [Layer N+1]          [Layer N+2]
+H2D Stream:      [H2D Shard N+1]   [H2D Shard N+2]
+AllGather:       [AG N+1]          [AG N+2]
+Slot Usage:      Slot0: Layer N    Slot1: Layer N+1
+```
+
+### Usage
+
+**CLI:**
+```bash
+# 4× GPU/NPU with AllGather (recommended)
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4
+
+# Without AllGather (each rank loads full weights, no sharding)
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4 \
+  --dlo-no-use-allgather
+
+# With SP instead of DP (long sequences)
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --usp 4
+```
+
+**Python API:**
+```python
+from vllm_omni import Omni
+
+m = Omni(
+    model="/path/to/model",
+    enable_distributed_layerwise_offload=True,
+    dlo_use_allgather=True,  # default
+)
+```
+
+### CLI Flags
+
+| Flag | Description | Default |
+|------|-------------|---------|
+| `--enable-distributed-layerwise-offload` | Enable DLO | `false` |
+| `--data-parallel-size N` | Number of DP ranks for weight sharding | `1` |
+| `--dlo-use-allgather` | Shard + AllGather (saves CPU, requires concurrent requests) | `true` |
+| `--dlo-no-use-allgather` | Full weights per rank (no sharding, no AllGather) | `false` |
+
+### How model weights are loaded (mmap path)
+
+When DLO + AllGather is active, the offloader:
+
+1. **Saves non-persistent buffers** (e.g. RoPE `inv_freq`, timestep `freqs`)
+   from the normally-created transformer
+2. **Converts to meta device** via `to_empty(device="meta")`, releasing random
+   initialization weights
+3. **Loads checkpoint weights as mmap views** via `safe_open().get_tensor()`,
+   which return views into the OS page cache (shared across all ranks, 0 RSS)
+4. **Calls `post_load_weights()`** to apply model-specific dtype conversions
+   (e.g. Cosmos3's `time_embedder` → FP32)
+5. **Restores non-persistent buffers** from saved copies
+
+This approach requires **zero model-specific code changes** — no pipeline or
+transformer modifications are needed.
+
+### OffloadPlan (declarative topology metadata)
+
+Models can optionally declare an `OffloadPlan` class variable to provide
+topology metadata (block attribute names, submodules to offload) without
+any offload-specific logic:
+
+```python
+from vllm_omni.diffusion.offloader import OffloadPlan
+
+class MyPipeline(nn.Module):
+    _dit_modules = ["transformer"]
+    _offload_plan = OffloadPlan(
+        block_attrs={"transformer": ("blocks",)},
+        offload_submodules={"context_encoder": "layers"},
+    )
+```
+
+When not declared, the offloader falls back to `_layerwise_offload_blocks_attrs`
+and heuristic attribute search.  This is backward-compatible — existing models
+work without any changes.
+
+### DP Multi-concurrency
+
+When `--data-parallel-size > 1` and AllGather is enabled, the scheduler batches
+up to `dp_size` requests per denoise step.  Each DP rank processes a different
+request while AllGather synchronizes only weight shards (request-independent).
+
+**Requirements for concurrent requests:**
+- `num_inference_steps` must be specified explicitly (None is not allowed)
+- All concurrent requests must have the same `num_inference_steps` value
+- These constraints exist because AllGather is a collective that requires
+  every rank to participate at each denoise step
+
+### Limitations
+- Online quantization (FP8) is incompatible with mmap loading — falls back to
+  regular `load_weights()` automatically
+- Tensor Parallel is not supported (DLO uses DP-based sharding)
+- HSDP + AllGather is rejected (would double-shard weights)
+- `num_inference_steps=None` is not allowed in DP multi-concurrency mode
 
 **Module Discovery**
 
@@ -208,7 +335,10 @@ OffloadBackend (base class)
 ├── ModelLevelOffloadBackend → uses SequentialOffloadHook (.to() swap)
 │                              (delegates to a pipeline's enable_omni_model_cpu_offload
 │                               for split models like Cosmos3)
-└── LayerWiseOffloadBackend → uses LayerwiseOffloadHook
+├── LayerWiseOffloadBackend → uses LayerwiseOffloadHook
+│                          (single-GPU, full weights on host)
+└── DistributedLayerwiseOffloadBackend → uses DistributedLayerwiseOffloadHook
+                                         (multi-GPU, 1/dp_size sharded weights + AllGather)
 ```
 
 Factory function `get_offload_backend()` selects the appropriate backend based on
@@ -221,20 +351,22 @@ reasoner/generator components inside the model forward pass.
 
 ## Supported Models
 
-| Architecture | Example Models | DiT Class | Model-Level Offload | Layerwise Offload | Blocks Attrs (Layerwise specific) |
-|--------------|----------------|-----------|---------------------|-------------------|-----------------------------------|
-| Flux2Pipeline | `black-forest-labs/FLUX.2-dev` | `Flux2Transformer2DModel` | ✓ | ✓ | `"transformer_blocks"`, `"single_transformer_blocks"` |
-| LongCatImagePipeline | `meituan-longcat/LongCat-Image` | `LongCatImageTransformer2DModel` | - | ✓ | `"transformer_blocks"`, `"single_transformer_blocks"` |
-| NextStep11Pipeline | `stepfun-ai/NextStep-1.1` | `NextStepModel` | - | ✓ | `"layers"` |
-| OvisImagePipeline | `AIDC-AI/Ovis-Image-7B` | `OvisImageTransformer2DModel` | - | ✓ | `"transformer"` |
-| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | ✓ | ✓ | `"transformer_blocks"` |
-| StableDiffusionXLPipeline | `stabilityai/stable-diffusion-xl-base-1.0` | `SDXLUNet2DConditionModel` | ✓ | ✓ | `"down_blocks"`, `"up_blocks"` |
-| StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | - | ✓ | `"transformer_blocks"` |
-| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | `"blocks"` |
-| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | `"blocks"` |
-| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` (`cfm_decoder.model.diff_estimator`) | ✓ | ✓ | `"layers"` |
-| BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | - | ✓ | `"layers"`, `"customized modules"` |
+| Architecture | Example Models | DiT Class | Model-Level Offload | Layerwise Offload | Distributed Layerwise Offload | Blocks Attrs (Layerwise specific) |
+|--------------|----------------|-----------|---------------------|-------------------|-------------------------------|-----------------------------------|
+| Flux2Pipeline | `black-forest-labs/FLUX.2-dev` | `Flux2Transformer2DModel` | ✓ | ✓ | - | `"transformer_blocks"`, `"single_transformer_blocks"` |
+| LongCatImagePipeline | `meituan-longcat/LongCat-Image` | `LongCatImageTransformer2DModel` | - | ✓ | - | `"transformer_blocks"`, `"single_transformer_blocks"` |
+| NextStep11Pipeline | `stepfun-ai/NextStep-1.1` | `NextStepModel` | - | ✓ | - | `"layers"` |
+| OvisImagePipeline | `AIDC-AI/Ovis-Image-7B` | `OvisImageTransformer2DModel` | - | ✓ | - | `"transformer"` |
+| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | ✓ | ✓ | - | `"transformer_blocks"` |
+| StableDiffusionXLPipeline | `stabilityai/stable-diffusion-xl-base-1.0` | `SDXLUNet2DConditionModel` | ✓ | ✓ | - | `"down_blocks"`, `"up_blocks"` |
+| StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | - | ✓ | - | `"transformer_blocks"` |
+| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | - | `"blocks"` |
+| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | - | `"blocks"` |
+| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` (`cfm_decoder.model.diff_estimator`) | ✓ | ✓ | - | `"layers"` |
+| BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | - | ✓ | - | `"layers"`, `"customized modules"` |
+| Cosmos3OmniDiffusersPipeline | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | `Cosmos3VFMTransformer`, `Cosmos3LanguageModel` | ✓ | ✓ | ✓ | `"layers"`, `"gen_layers"` |
 
 **Notes:**
 - Model-Level Offloading is expected to be supported by all common diffusion models (DiT and encoders) naturally
 - Layerwise Offloading requires DiT class to define `_layerwise_offload_blocks_attrs` pointing to transformer blocks
+- Distributed Layerwise Offloading works with any model that supports Layerwise Offloading — no additional model changes required.  See [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md) for usage examples.

docs/user_guide/examples/online_serving/bagel.md

@@ -40,7 +40,7 @@ bash run_server_stage_cli.sh --stage 0
 bash run_server_stage_cli.sh --stage 1
 ```
 
-To use a custom deploy YAML (note: `--stage-configs-path` is deprecated in favor of `--deploy-config`):
+To use a custom deploy YAML:
 
 ```bash
 vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 \

docs/user_guide/examples/online_serving/glm_image.md

@@ -11,8 +11,8 @@ memory, sampling params) live in `vllm_omni/deploy/glm_image.yaml`.
 vllm serve zai-org/GLM-Image --omni --port 8091
 ```
 
-The config system auto-detects the pipeline from the model's `model_index.json` — no
-manual `--stage-configs-path` or `--deploy-config` needed.
+The config system auto-detects the pipeline from the model's `model_index.json`, so
+no manual `--deploy-config` is needed.
 
 By default, stage 0 (AR) runs on GPU 0 and stage 1 (Diffusion) on GPU 1. To colocate
 both stages on a single GPU, override per stage:

docs/user_guide/examples/online_serving/qwen2_5_omni.md

@@ -15,9 +15,9 @@ Please refer to [README.md](https://github.com/vllm-project/vllm-omni/tree/main/
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091
 ```
 
-If you have custom stage configs file, launch the server with command below
+To use a custom deploy config, launch the server with the command below:
 ```bash
-vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
+vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --deploy-config /path/to/deploy_config.yaml
 ```
 
 ### Send Multi-modal Request
@@ -182,7 +182,7 @@ The script supports the following arguments:
 - `--model`: Model name/path (default: Qwen/Qwen2.5-Omni-7B)
 - `--server-port`: Port for vLLM server (default: 8091)
 - `--gradio-port`: Port for Gradio demo (default: 7861)
-- `--stage-configs-path`: Path to custom stage configs YAML file (optional)
+- `--deploy-config`: Path to a custom deploy config YAML file (optional)
 - `--server-host`: Host for vLLM server (default: 0.0.0.0)
 - `--gradio-ip`: IP for Gradio demo (default: 127.0.0.1)
 - `--share`: Share Gradio demo publicly (creates a public link)
@@ -195,9 +195,9 @@ The script supports the following arguments:
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091
 ```
 
-If you have custom stage configs file:
+To use a custom deploy config:
 ```bash
-vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
+vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --deploy-config /path/to/deploy_config.yaml
 ```
 
 **Step 2: Run the Gradio demo**

docs/user_guide/examples/online_serving/qwen3_omni.md

@@ -306,7 +306,7 @@ The gradio script supports the following arguments:
     ``````
 ??? abstract "qwen3_omni_moe_thinking.yaml"
     ``````yaml
-    --8<-- "examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml"
+    --8<-- "vllm_omni/deploy/qwen3_omni_moe_thinking.yaml"
     ``````
 ??? abstract "run_curl_multimodal_generation.sh"
     ``````sh

docs/user_guide/examples/online_serving/text_to_speech.md

@@ -422,12 +422,19 @@ curl -X POST http://localhost:8091/v1/audio/speech \
 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:
+The `/v1/audio/speech/stream` endpoint accepts text incrementally and synthesizes the buffered text as one continuous request on `input.done`:
 ```bash
 python qwen3_tts/streaming_speech_client.py --text "Hello world. How are you? I am fine."
 python qwen3_tts/streaming_speech_client.py --text "..." --simulate-stt --stt-delay 0.1
 ```
 
+`input.done` flushes without closing, so repeating `--text` synthesizes several utterances over one connection:
+```bash
+python qwen3_tts/streaming_speech_client.py \
+    --text "First utterance." \
+    --text "Second utterance, same connection."
+```
+
 ### Gradio demos
 ```bash
 ./qwen3_tts/run_gradio_demo.sh                              # CustomVoice (default)

docs/user_guide/fault_injection_reliability_matrix.md

@@ -69,19 +69,19 @@ The behaviors and conclusions above are summarized from current fault injection
 
 ### Example: `SIGTERM` fault injection log (Qwen3-Omni)
 
-The following full log is from a real run where a `SIGTERM` signal was sent to the serve root process.
+The following log is based on a real run where a `SIGTERM` signal was sent to the serve root process. Its command line has been normalized to the current `--deploy-config` option.
 
 ```text
-Launching OmniServer with: /workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --stage-configs-path vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
+Launching OmniServer with: /workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --deploy-config vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
 Server ready on 127.0.0.1:60675 (OmniServer startup took 206.023s)
 OmniServer started successfully
-[reliability][process-kill] current_server_proc pid=556855 name=python3 cmdline=/workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --stage-configs-path vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
+[reliability][process-kill] current_server_proc pid=556855 name=python3 cmdline=/workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --deploy-config vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
 [reliability][process-kill] current_server_proc pid=557118 name=python3 cmdline=/workspace/.venv/bin/python3 -c from multiprocessing.resource_tracker import main;main(57)
 [reliability][process-kill] current_server_proc pid=557119 name=VLLM::StageEngineCoreProc_noid_replica0_DP0 cmdline=VLLM::StageEngineCoreProc_noid_replica0_DP0
 [reliability][process-kill] current_server_proc pid=557122 name=VLLM::StageEngineCoreProc_noid_replica0_DP0 cmdline=VLLM::StageEngineCoreProc_noid_replica0_DP0
 [reliability][process-kill] current_server_proc pid=558201 name=VLLM::StageEngineCoreProc_noid_replica0_DP0 cmdline=VLLM::StageEngineCoreProc_noid_replica0_DP0
 
-[reliability][process-kill] root-kill pid=556855 name=python3 signal=SIGTERM cmdline=/workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --stage-configs-path vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
+[reliability][process-kill] root-kill pid=556855 name=python3 signal=SIGTERM cmdline=/workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --deploy-config vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
 [0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [omni_base.py:463] [AsyncOmni] Shutting down
 [0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [async_omni_engine.py:2397] [AsyncOmniEngine] Shutting down Orchestrator
 [0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [orchestrator.py:351] [Orchestrator] Received shutdown signal

docs/user_guide/quantization/fp8.md

@@ -93,6 +93,7 @@ warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])
 | HunyuanImage-3.0 | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | Yes | Yes | All layers; use the Hunyuan stage config for multi-stage runs | None | |
 | HunyuanVideo-1.5 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `720p_t2v`, `480p_i2v` | Yes | Yes | All layers | None | |
 | Cosmos3 | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | Yes | Not validated | All layers | None | |
+| MiniMax H3 | `MiniMaxAI/MiniMax-H3` | Yes | Not validated | DiT linears except patch, timestep, and final projections; not compatible with layerwise offload | None | |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 

examples/model_display_names.yml

@@ -8,5 +8,4 @@ mimo_audio: MiMo-Audio
 ming_flash_omni: Ming-flash-omni 2.0
 qwen2_5_omni: Qwen2.5-Omni
 qwen3_omni: Qwen3-Omni
-sensenova_u1: SenseNova-U1
 step_audio2: Step-Audio2

examples/offline_inference/image_to_video/image_to_video.py

@@ -334,6 +334,7 @@ def main():
     is_ltx23 = "ltx23" in model_class_name_lower or "ltx-2.3" in model_name
     is_ltx2 = is_ltx2_distilled or is_ltx23 or "ltx2" in model_class_name_lower or "ltx-2" in model_name
     is_cosmos = "cosmos" in model_name or (model_class_name is not None and "cosmos" in model_class_name.lower())
+    is_cosmos_edge = is_cosmos and ("edge" in model_name or "edge" in model_class_name_lower)
 
     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
@@ -347,7 +348,19 @@ def main():
 
     # Per-model generation defaults, applied only when the matching flag is omitted.
     # Cosmos3 would otherwise silently inherit the Wan2.2 defaults (wrong size/steps/shift).
-    if is_cosmos:
+    if is_cosmos_edge:
+        # Cosmos3-Edge native defaults: 480x832, gs 5.0, flow_shift 3.0 — NOT the Nano/Super
+        # 720p / gs 6.0 / flow_shift 10.0 below, which produce degenerate output on Edge.
+        d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = (
+            24,
+            5.0,
+            189,
+            35,
+            3.0,
+            480 * 832,
+            16,
+        )
+    elif is_cosmos:
         d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = (
             24,
             6.0,

examples/offline_inference/text_to_speech/cosyvoice3/end2end.py

@@ -182,10 +182,13 @@ def run_e2e():
                 print(f"Multimodal output keys: {mm.keys()}")
                 if "audio" in mm:
                     audio_out = mm["audio"]
+                    # The model emits its own rate (24 kHz for CosyVoice3); take
+                    # it from the payload so the file is not mislabeled.
+                    sample_rate = int(mm["sr"]) if "sr" in mm.keys() else 24000
                     print(f"Generated Audio Shape: {audio_out.shape}")
                     out_path = f"output_{i}.wav"
-                    sf.write(out_path, audio_out.cpu().numpy().squeeze(), 22050)
-                    print(f"Saved audio to {out_path}")
+                    sf.write(out_path, audio_out.cpu().numpy().squeeze(), sample_rate)
+                    print(f"Saved audio to {out_path} (sr={sample_rate})")
             else:
                 print("No multimodal output found.")
         except Exception as e:

examples/offline_inference/text_to_video/text_to_video.py

@@ -56,6 +56,19 @@ _MODEL_PRESETS = {
         "flow_shift": 10.0,
         "output": "cosmos3_t2v_output.mp4",
     },
+    # Cosmos3-Edge is a compact checkpoint with its own native resolution / guidance /
+    # flow_shift. It must NOT inherit the Nano/Super "cosmos" numbers above (720p / gs 6.0 /
+    # flow_shift 10.0), which produce degenerate output on Edge.
+    "cosmos3_edge": {
+        "height": 480,
+        "width": 832,
+        "num_frames": 189,
+        "num_inference_steps": 35,
+        "guidance_scale": 5.0,
+        "fps": 24,
+        "flow_shift": 3.0,
+        "output": "cosmos3_edge_t2v_output.mp4",
+    },
     "helios": {
         "height": 384,
         "width": 640,
@@ -103,6 +116,10 @@ def _detect_preset(model: str, model_class_name: str | None = None) -> dict:
         return _MODEL_PRESETS["ltx2"]
     if "vace" in model_lower:
         return _MODEL_PRESETS["vace"]
+    # Edge must be matched before the generic cosmos branch (its "cosmos" substring would
+    # otherwise pick up the Nano/Super 720p / gs 6.0 / flow_shift 10.0 preset).
+    if ("cosmos" in model_lower or "cosmos" in class_lower) and ("edge" in model_lower or "edge" in class_lower):
+        return _MODEL_PRESETS["cosmos3_edge"]
     if "cosmos" in model_lower:
         return _MODEL_PRESETS["cosmos"]
     if "hunyuan" in model_lower:

examples/online_serving/audex/run_server.sh

@@ -53,5 +53,5 @@ vllm-omni serve "$MODEL" \
     --host 0.0.0.0 \
     --port "$PORT" \
     --trust-remote-code \
-    --stage-configs-path "$DEPLOY_YAML" \
+    --deploy-config "$DEPLOY_YAML" \
     --omni

examples/online_serving/chart-helm/templates/_helpers.tpl

@@ -28,8 +28,11 @@ If image.command is set, uses that as a full override.
 - {{ .Values.omniArgs.numGpus | quote }}
 {{- end }}
 {{- if .Values.omniArgs.stageConfigsPath }}
-- "--stage-configs-path"
-- {{ .Values.omniArgs.stageConfigsPath | quote }}
+{{- fail "omniArgs.stageConfigsPath has been removed; use omniArgs.deployConfigPath" }}
+{{- end }}
+{{- if .Values.omniArgs.deployConfigPath }}
+- "--deploy-config"
+- {{ .Values.omniArgs.deployConfigPath | quote }}
 {{- end }}
 {{- if and .Values.omniArgs.cacheBackend (ne .Values.omniArgs.cacheBackend "none") }}
 - "--cache-backend"

examples/online_serving/chart-helm/tests/deployment_test.yaml

@@ -78,6 +78,26 @@ tests:
           path: spec.template.spec.containers[0].command
           content: "--enable-layerwise-offload"
 
+  - it: should include deploy config in command
+    set:
+      omniArgs:
+        deployConfigPath: "/configs/deploy.yaml"
+    asserts:
+      - equal:
+          path: spec.template.spec.containers[0].command[8]
+          value: "--deploy-config"
+      - equal:
+          path: spec.template.spec.containers[0].command[9]
+          value: "/configs/deploy.yaml"
+
+  - it: should reject removed stage config value
+    set:
+      omniArgs:
+        stageConfigsPath: "/configs/legacy.yaml"
+    asserts:
+      - failedTemplate:
+          errorMessage: "omniArgs.stageConfigsPath has been removed; use omniArgs.deployConfigPath"
+
   - it: should use full command override when image.command is set
     set:
       image:

examples/online_serving/chart-helm/values.yaml

@@ -67,8 +67,8 @@ omniArgs:
   enableCpuOffload: false
   # -- Number of GPUs for diffusion inference (empty = auto)
   numGpus:
-  # -- Path to stage configs file (empty = auto-detected from model)
-  stageConfigsPath:
+  # -- Path to deploy config file (empty = auto-detected from model)
+  deployConfigPath:
   # -- Cache backend: none, tea_cache, or cache_dit
   cacheBackend: "none"
   # -- Default sampling params as a JSON string

examples/online_serving/minicpmo/README.md

@@ -23,17 +23,18 @@ including `librosa`.
 
 The deploy config auto-loads via `--omni`.
 The default `vllm_omni/deploy/minicpmo_4_5.yaml` keeps all three stages on
-logical device 0 with memory budgets of 55%, 22%, and 22%. The profile admits
-at most four sequences per stage and bounds startup video profiling to 32
-frames per video. For throughput,`minicpmo_4_5_batching.yaml` gives the Thinker
+logical device 0 with memory budgets of 55%, 15%, and 15%, leaving headroom
+for runtime kernels such as the HiFi-GAN vocoder's cuDNN workspace. The
+profile admits at most four sequences per stage and bounds startup video
+profiling to 32 frames per video. For throughput,
+`minicpmo_4_5_2gpu.yaml` gives the Thinker
 GPU 0 (90%) and colocates the Talker (55%) and Code2Wav (35%) on GPU 1. That
 profile admits at most four concurrent sequences per stage.
 
 | deploy config | GPUs | Notes |
 |---|---|---|
 | `minicpmo_4_5.yaml` (default) | 1 | Memory-constrained compatibility layout. |
-| `minicpmo_4_5_batching.yaml` | 2 | Recommended continuous-batching layout; Talker and Code2Wav share GPU 1. |
-| `minicpmo_4_5_2gpu.yaml` | 2 | Backward-compatible alias for `minicpmo_4_5_batching.yaml`. |
+| `minicpmo_4_5_2gpu.yaml` | 2 | Recommended continuous-batching layout; Talker and Code2Wav share GPU 1. |
 | `minicpmo_4_5_3gpu.yaml` | 3 | One GPU per stage. |
 | `minicpmo_4_5_8x4090.yaml` | 8 | Full 8x4090 layout. |
 | `minicpmo_4_5_duplex.yaml` | 1 | Experimental native full-duplex overlay. |
@@ -89,17 +90,18 @@ python openai_chat_completion_client_for_multimodal_generation.py \
     --prompt "Briefly introduce yourself."
 ```
 
-Streaming text + audio:
+Streaming text + audio (use `--stream`):
 
 ```bash
-python streaming_chat_completion.py \
-    --base-url http://localhost:8099/v1 \
-    --output minicpmo_stream.wav
+python openai_chat_completion_client_for_multimodal_generation.py \
+    --query-type text \
+    --prompt "Briefly introduce yourself." \
+    --port 8099 \
+    --stream
 ```
 
-The example prints text deltas immediately and joins the independently encoded
-24 kHz WAV audio deltas into one valid WAV file. Add `--text-only` to skip
-Talker and Code2Wav.
+The client prints text deltas as they arrive and saves streamed audio chunks
+to WAV files.
 
 Shared helpers also work if you pass MiniCPM defaults yourself:
 

examples/online_serving/minicpmo/realtime_web/app/static/playback_worklet.js

@@ -6,6 +6,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
     this.playedFrames = 0;
     this.underrunFrames = 0;
     this.drain = null;
+    this.terminalFadeFrames = 0;
     this.started = false;
     this.activeResponseId = null;
     this.initialBufferFrames = Math.round(sampleRate * 0.2);
@@ -31,6 +32,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
       }
     } else if (message.type === 'drain') {
       this.drain = { responseId: message.responseId || null };
+      this.terminalFadeFrames = Math.min(this.fadeFrames, this.bufferedFrames());
       if (!this.started && this.bufferedFrames() > 0) {
         this.bufferWaitFrames = 0;
         this.startPlayback();
@@ -42,6 +44,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
       this.playedFrames = 0;
       this.underrunFrames = 0;
       this.drain = null;
+      this.terminalFadeFrames = 0;
       this.started = false;
       this.activeResponseId = null;
       this.bufferWaitFrames = this.initialBufferFrames;
@@ -76,6 +79,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
     this.playedFrames = 0;
     this.underrunFrames = 0;
     this.drain = null;
+    this.terminalFadeFrames = 0;
     this.started = false;
     this.activeResponseId = null;
     this.bufferWaitFrames = this.initialBufferFrames;
@@ -116,6 +120,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
     while (target < output.length && this.queue.length > 0) {
       const pcm = this.queue[0];
       const count = Math.min(output.length - target, pcm.length - this.offset);
+      const remainingBeforeChunk = this.bufferedFrames();
       for (let index = 0; index < count; index += 1) {
         let sample = pcm[this.offset + index] / 32768;
         if (this.fadeInFrames > 0) {
@@ -123,6 +128,16 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
           sample *= elapsed / this.fadeFrames;
           this.fadeInFrames -= 1;
         }
+        const remainingFrames = remainingBeforeChunk - index;
+        if (
+          this.drain
+          && this.terminalFadeFrames > 0
+          && remainingFrames <= this.terminalFadeFrames
+        ) {
+          sample *= this.terminalFadeFrames === 1
+            ? 0
+            : (remainingFrames - 1) / (this.terminalFadeFrames - 1);
+        }
         output[target + index] = sample;
       }
       target += count;

examples/online_serving/minicpmo/run_gradio_demo.sh

@@ -3,7 +3,7 @@
 #
 # Prereq:
 #   Start a vllm-omni OpenAI server for MiniCPM-o 4.5 on :8099 (see the
-#   8x4090 stage config under vllm_omni/model_executor/stage_configs).
+#   8x4090 deploy config at vllm_omni/deploy/minicpmo_4_5_8x4090.yaml).
 set -e
 
 HERE="$(cd "$(dirname "$0")" && pwd)"

examples/online_serving/minicpmo/streaming_chat_completion.py

@@ -1,71 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Stream MiniCPM-o 4.5 text and audio from /v1/chat/completions."""
-
-from __future__ import annotations
-
-import argparse
-import base64
-import io
-import wave
-from pathlib import Path
-
-from openai import OpenAI
-
-
-def append_wav_chunk(encoded: str, pcm_parts: list[bytes]) -> tuple[int, int, int]:
-    """Decode one base64 WAV delta and append its PCM frames."""
-    with wave.open(io.BytesIO(base64.b64decode(encoded)), "rb") as wav:
-        channels = wav.getnchannels()
-        sample_width = wav.getsampwidth()
-        sample_rate = wav.getframerate()
-        pcm_parts.append(wav.readframes(wav.getnframes()))
-    return channels, sample_width, sample_rate
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--base-url", default="http://localhost:8099/v1")
-    parser.add_argument("--model", default="openbmb/MiniCPM-o-4_5")
-    parser.add_argument("--prompt", default="Say hello, then introduce vLLM in one sentence.")
-    parser.add_argument("--output", type=Path, default=Path("minicpmo_stream.wav"))
-    parser.add_argument("--text-only", action="store_true")
-    args = parser.parse_args()
-
-    modalities = ["text"] if args.text_only else ["text", "audio"]
-    stream = OpenAI(base_url=args.base_url, api_key="EMPTY").chat.completions.create(
-        model=args.model,
-        messages=[{"role": "user", "content": args.prompt}],
-        modalities=modalities,
-        stream=True,
-        extra_body={"chat_template_kwargs": {"use_tts_template": not args.text_only}},
-    )
-
-    pcm_parts: list[bytes] = []
-    audio_format: tuple[int, int, int] | None = None
-    for chunk in stream:
-        for choice in chunk.choices:
-            content = getattr(choice.delta, "content", None)
-            if not content:
-                continue
-            if getattr(chunk, "modality", None) == "audio":
-                current_format = append_wav_chunk(content, pcm_parts)
-                if audio_format is not None and current_format != audio_format:
-                    raise ValueError(f"Audio format changed: {audio_format} -> {current_format}")
-                audio_format = current_format
-            elif getattr(chunk, "modality", None) == "text":
-                print(content, end="", flush=True)
-
-    print()
-    if pcm_parts and audio_format is not None:
-        channels, sample_width, sample_rate = audio_format
-        with wave.open(str(args.output), "wb") as wav:
-            wav.setnchannels(channels)
-            wav.setsampwidth(sample_width)
-            wav.setframerate(sample_rate)
-            wav.writeframes(b"".join(pcm_parts))
-        print(f"Audio saved to {args.output} ({sample_rate} Hz)")
-
-
-if __name__ == "__main__":
-    main()

examples/online_serving/qwen2_5_omni/README.md

@@ -12,9 +12,9 @@ Please refer to [README.md](../../../README.md)
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091
 ```
 
-If you have custom stage configs file, launch the server with command below
+To use a custom deploy config, launch the server with the command below:
 ```bash
-vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
+vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --deploy-config /path/to/deploy_config.yaml
 ```
 
 ### Send Multi-modal Request
@@ -173,7 +173,7 @@ The script supports the following arguments:
 - `--model`: Model name/path (default: Qwen/Qwen2.5-Omni-7B)
 - `--server-port`: Port for vLLM server (default: 8091)
 - `--gradio-port`: Port for Gradio demo (default: 7861)
-- `--stage-configs-path`: Path to custom stage configs YAML file (optional)
+- `--deploy-config`: Path to a custom deploy config YAML file (optional)
 - `--server-host`: Host for vLLM server (default: 0.0.0.0)
 - `--gradio-ip`: IP for Gradio demo (default: 127.0.0.1)
 - `--share`: Share Gradio demo publicly (creates a public link)
@@ -186,9 +186,9 @@ The script supports the following arguments:
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091
 ```
 
-If you have custom stage configs file:
+To use a custom deploy config:
 ```bash
-vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
+vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --deploy-config /path/to/deploy_config.yaml
 ```
 
 **Step 2: Run the Gradio demo**

examples/online_serving/qwen2_5_omni/run_gradio_demo.sh

@@ -13,7 +13,7 @@ set -e
 MODEL="Qwen/Qwen2.5-Omni-7B"
 SERVER_PORT=8091
 GRADIO_PORT=7861
-STAGE_CONFIGS_PATH=""
+DEPLOY_CONFIG_PATH=""
 SERVER_HOST="0.0.0.0"
 GRADIO_IP="127.0.0.1"
 GRADIO_SHARE=false
@@ -33,8 +33,8 @@ while [[ $# -gt 0 ]]; do
             GRADIO_PORT="$2"
             shift 2
             ;;
-        --stage-configs-path)
-            STAGE_CONFIGS_PATH="$2"
+        --deploy-config)
+            DEPLOY_CONFIG_PATH="$2"
             shift 2
             ;;
         --server-host)
@@ -56,7 +56,7 @@ while [[ $# -gt 0 ]]; do
             echo "  --model MODEL                 Model name/path (default: Qwen/Qwen2.5-Omni-7B)"
             echo "  --server-port PORT            Port for vLLM server (default: 8091)"
             echo "  --gradio-port PORT            Port for Gradio demo (default: 7861)"
-            echo "  --stage-configs-path PATH     Path to custom stage configs YAML file (optional)"
+            echo "  --deploy-config PATH          Path to custom deploy config YAML file (optional)"
             echo "  --server-host HOST            Host for vLLM server (default: 0.0.0.0)"
             echo "  --gradio-ip IP                IP for Gradio demo (default: 127.0.0.1)"
             echo "  --share                       Share Gradio demo publicly"
@@ -87,8 +87,8 @@ echo "=========================================="
 
 # Build vLLM server command
 SERVER_CMD=("vllm" "serve" "$MODEL" "--omni" "--port" "$SERVER_PORT" "--host" "$SERVER_HOST")
-if [ -n "$STAGE_CONFIGS_PATH" ]; then
-    SERVER_CMD+=("--stage-configs-path" "$STAGE_CONFIGS_PATH")
+if [ -n "$DEPLOY_CONFIG_PATH" ]; then
+    SERVER_CMD+=("--deploy-config" "$DEPLOY_CONFIG_PATH")
 fi
 
 # Function to cleanup on exit

examples/online_serving/qwen3_omni/qwen3_omni_moe_thinking.yaml

@@ -1,36 +0,0 @@
-# Stage config for running Qwen3-Omni-MoE-Thinking (text-only output)
-# This config is for models like Qwen3-Omni-30B-A3B-Thinking that only have the
-# thinker component and do not support audio output.
-#
-# Single stage: Thinker (multimodal understanding + text generation)
-
-# The following config has been verified on 2x H100-80G GPUs.
-stage_args:
-  - stage_id: 0
-    runtime:
-      devices: "0,1"
-    engine_args:
-      model_stage: thinker
-      max_num_seqs: 1
-      model_arch: Qwen3OmniMoeForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.9
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: text
-      distributed_executor_backend: "mp"
-      enable_prefix_caching: true
-      hf_config_name: thinker_config
-      tensor_parallel_size: 2
-    final_output: true
-    final_output_type: text
-    is_comprehension: true
-    default_sampling_params:
-      temperature: 0.4
-      top_p: 0.9
-      top_k: 1
-      max_tokens: 2048
-      seed: 42
-      detokenize: True
-      repetition_penalty: 1.05

examples/online_serving/qwen3_omni/run_gradio_demo.sh

@@ -13,7 +13,7 @@ set -e
 MODEL="Qwen/Qwen3-Omni-30B-A3B-Instruct"
 SERVER_PORT=8091
 GRADIO_PORT=7861
-STAGE_CONFIGS_PATH=""
+DEPLOY_CONFIG=""
 SERVER_HOST="0.0.0.0"
 GRADIO_IP="127.0.0.1"
 GRADIO_SHARE=false
@@ -33,8 +33,8 @@ while [[ $# -gt 0 ]]; do
             GRADIO_PORT="$2"
             shift 2
             ;;
-        --stage-configs-path)
-            STAGE_CONFIGS_PATH="$2"
+        --deploy-config)
+            DEPLOY_CONFIG="$2"
             shift 2
             ;;
         --server-host)
@@ -56,7 +56,7 @@ while [[ $# -gt 0 ]]; do
             echo "  --model MODEL                 Model name/path (default: Qwen/Qwen3-Omni-30B-A3B-Instruct)"
             echo "  --server-port PORT            Port for vLLM server (default: 8091)"
             echo "  --gradio-port PORT            Port for Gradio demo (default: 7861)"
-            echo "  --stage-configs-path PATH     Path to custom stage configs YAML file (optional)"
+            echo "  --deploy-config PATH          Path to custom deploy config YAML file (optional)"
             echo "  --server-host HOST            Host for vLLM server (default: 0.0.0.0)"
             echo "  --gradio-ip IP                IP for Gradio demo (default: 127.0.0.1)"
             echo "  --share                       Share Gradio demo publicly"
@@ -87,8 +87,8 @@ echo "=========================================="
 
 # Build vLLM server command
 SERVER_CMD=("vllm" "serve" "$MODEL" "--omni" "--port" "$SERVER_PORT" "--host" "$SERVER_HOST")
-if [ -n "$STAGE_CONFIGS_PATH" ]; then
-    SERVER_CMD+=("--stage-configs-path" "$STAGE_CONFIGS_PATH")
+if [ -n "$DEPLOY_CONFIG" ]; then
+    SERVER_CMD+=("--deploy-config" "$DEPLOY_CONFIG")
 fi
 
 # Function to cleanup on exit

examples/online_serving/text_to_speech/README.md

@@ -336,7 +336,7 @@ vllm serve OpenMOSS-Team/MOSS-TTS-Nano --omni --port 8091
 # or:
 ./moss_tts_nano/run_server.sh
 ```
-The deploy config at `vllm_omni/deploy/moss_tts_nano.yaml` auto-loads; no `--stage-configs-path`, `--trust-remote-code`, or `--enforce-eager` flags are needed.
+The deploy config at `vllm_omni/deploy/moss_tts_nano.yaml` auto-loads; no `--deploy-config`, `--trust-remote-code`, or `--enforce-eager` flags are needed.
 
 ### Sending requests
 ```bash
@@ -537,12 +537,19 @@ curl -X POST http://localhost:8091/v1/audio/speech \
 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:
+The `/v1/audio/speech/stream` endpoint accepts text incrementally and synthesizes the buffered text as one continuous request on `input.done`:
 ```bash
 python qwen3_tts/streaming_speech_client.py --text "Hello world. How are you? I am fine."
 python qwen3_tts/streaming_speech_client.py --text "..." --simulate-stt --stt-delay 0.1
 ```
 
+`input.done` flushes without closing, so repeating `--text` synthesizes several utterances over one connection:
+```bash
+python qwen3_tts/streaming_speech_client.py \
+    --text "First utterance." \
+    --text "Second utterance, same connection."
+```
+
 To receive word-level timestamps, launch the server with a forced aligner:
 ```bash
 vllm-omni serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \

examples/online_serving/text_to_speech/qwen3_tts/streaming_speech_client.py

@@ -7,6 +7,12 @@ Usage:
     # Send full text at once
     python streaming_speech_client.py --text "Hello world. How are you? I am fine."
 
+    # Several utterances over one connection: input.done flushes each of them
+    # and the connection stays open, so only the first pays the handshake
+    python streaming_speech_client.py \
+        --text "First utterance." \
+        --text "Second utterance, same connection."
+
     # Pick a built-in speaker for CustomVoice models
     python streaming_speech_client.py \
         --text "打开电子枪，关闭扫描，切换到低倍模式。" \
@@ -64,18 +70,27 @@ except ImportError:
     raise SystemExit(1)
 
 
-def save_audio_file(output_dir: str, sentence_index: int, response_format: str, chunks: list[bytes]) -> str:
+def frame_basename(msg: dict) -> str:
+    """Name a file after the utterance and sentence a frame belongs to.
+
+    Every utterance is sentence 0, so the utterance index is what keeps the
+    files of one connection from overwriting each other.
+    """
+    return f"utterance_{msg['utterance_index']:03d}_sentence_{msg['sentence_index']:03d}"
+
+
+def save_audio_file(output_dir: str, basename: str, response_format: str, chunks: list[bytes]) -> str:
     audio_bytes = b"".join(chunks)
     if response_format == "pcm":
-        raw_path = os.path.join(output_dir, f"sentence_{sentence_index:03d}.pcm")
+        raw_path = os.path.join(output_dir, f"{basename}.pcm")
         with open(raw_path, "wb") as f:
             f.write(audio_bytes)
 
-        wav_path = os.path.join(output_dir, f"sentence_{sentence_index:03d}.wav")
+        wav_path = os.path.join(output_dir, f"{basename}.wav")
         save_pcm_wav(wav_path, audio_bytes)
         return wav_path
 
-    filename = os.path.join(output_dir, f"sentence_{sentence_index:03d}.{response_format}")
+    filename = os.path.join(output_dir, f"{basename}.{response_format}")
     with open(filename, "wb") as f:
         f.write(audio_bytes)
     return filename
@@ -95,183 +110,207 @@ def timestamp_words_text(timestamps: list[dict] | None) -> str:
     return " ".join(str(item.get("word", "")) for item in timestamps if item.get("word"))
 
 
+async def send_utterance(ws, text: str, simulate_stt: bool, stt_delay: float) -> None:
+    """Send one utterance's text, then input.done to flush it."""
+    if simulate_stt:
+        words = text.split(" ")
+        for i, word in enumerate(words):
+            chunk = word + (" " if i < len(words) - 1 else "")
+            await ws.send(
+                json.dumps(
+                    {
+                        "type": "input.text",
+                        "text": chunk,
+                    }
+                )
+            )
+            print(f"  Sent: {chunk!r}")
+            await asyncio.sleep(stt_delay)
+    else:
+        await ws.send(
+            json.dumps(
+                {
+                    "type": "input.text",
+                    "text": text,
+                }
+            )
+        )
+        print(f"Sent full text: {text!r}")
+
+    await ws.send(json.dumps({"type": "input.done"}))
+    print("Sent input.done")
+
+
+async def receive_utterance(
+    ws,
+    output_dir: str,
+    response_format: str,
+    word_timestamps: bool,
+    save_chunks: bool,
+) -> None:
+    """Consume frames until session.done marks the flushed utterance complete."""
+    current_label = "utterance 0"
+    current_sentence_text = ""
+    current_chunks: list[bytes] = []
+    current_timestamps: list[dict] = []
+    # Distinguish `null` (aligner failed) from `[]` (silence / no tokens):
+    # the sentence is a failure only if no frame ever carried timestamps.
+    alignment_frame_seen = False
+
+    while True:
+        message = await ws.recv()
+
+        if isinstance(message, bytes):
+            current_chunks.append(message)
+            print(f"  Received audio chunk for {current_label}: {len(message)} bytes")
+        else:
+            # JSON frame
+            msg = json.loads(message)
+            msg_type = msg.get("type")
+
+            if msg_type == "audio.start":
+                current_label = f"utterance {msg['utterance_index']}"
+                current_sentence_text = msg.get("sentence_text", "")
+                current_chunks = []
+                current_timestamps = []
+                alignment_frame_seen = False
+                print(f"  [{current_label}] Generating: {current_sentence_text!r}")
+            elif msg_type == "audio.chunk":
+                audio = base64.b64decode(msg["audio_b64"])
+                current_chunks.append(audio)
+                chunk_timestamps = msg.get("timestamps")
+                # Sentence-level alignment sends a trailing timestamp-only
+                # frame with empty audio; don't write an empty chunk file.
+                if save_chunks and audio:
+                    chunk_base = os.path.join(
+                        output_dir,
+                        f"{frame_basename(msg)}_chunk_{msg['chunk_id']:03d}",
+                    )
+                    chunk_pcm = f"{chunk_base}.pcm"
+                    chunk_wav = f"{chunk_base}.wav"
+                    chunk_json = f"{chunk_base}_timestamps.json"
+                    with open(chunk_pcm, "wb") as f:
+                        f.write(audio)
+                    save_pcm_wav(chunk_wav, audio, int(msg.get("sample_rate", 24000)))
+                    with open(chunk_json, "w", encoding="utf-8") as f:
+                        json.dump(
+                            {
+                                "utterance_index": msg["utterance_index"],
+                                "sentence_index": msg["sentence_index"],
+                                "chunk_id": msg["chunk_id"],
+                                "chunk_start_ms": msg.get("chunk_start_ms"),
+                                "chunk_end_ms": msg.get("chunk_end_ms"),
+                                "sentence_text": current_sentence_text,
+                                "covered_text": timestamp_words_text(chunk_timestamps),
+                                "timestamps": chunk_timestamps,
+                            },
+                            f,
+                            ensure_ascii=False,
+                            indent=2,
+                        )
+                if chunk_timestamps is None:
+                    print(f"  [{current_label}] chunk {msg['chunk_id']}: {len(audio)} bytes, timestamps unavailable")
+                else:
+                    alignment_frame_seen = True
+                    current_timestamps.extend(chunk_timestamps)
+                    print(
+                        f"  [{current_label}] chunk {msg['chunk_id']}: "
+                        f"{len(audio)} bytes, {len(chunk_timestamps)} timestamp(s)"
+                    )
+                if save_chunks and audio:
+                    print(f"    saved chunk -> {chunk_wav}")
+            elif msg_type == "audio.done":
+                filename = save_audio_file(
+                    output_dir,
+                    frame_basename(msg),
+                    response_format,
+                    current_chunks,
+                )
+                if word_timestamps:
+                    # `null` on failure, else the accumulated list (`[]` for silence).
+                    sentence_timestamps = current_timestamps if alignment_frame_seen else None
+                    ts_filename = os.path.join(
+                        output_dir,
+                        f"{frame_basename(msg)}_timestamps.json",
+                    )
+                    with open(ts_filename, "w", encoding="utf-8") as f:
+                        json.dump(
+                            {
+                                "utterance_index": msg["utterance_index"],
+                                "sentence_index": msg["sentence_index"],
+                                "sentence_text": current_sentence_text,
+                                "covered_text": timestamp_words_text(sentence_timestamps),
+                                "timestamps": sentence_timestamps,
+                            },
+                            f,
+                            ensure_ascii=False,
+                            indent=2,
+                        )
+                print(
+                    f"  [{current_label}] Done"
+                    f" bytes={msg.get('total_bytes', len(b''.join(current_chunks)))}"
+                    f" error={msg.get('error', False)}"
+                    f" -> {filename}"
+                )
+                if word_timestamps:
+                    print(f"  [{current_label}] Timestamps -> {ts_filename}")
+                current_chunks = []
+                current_timestamps = []
+                alignment_frame_seen = False
+            elif msg_type == "session.done":
+                print(f"\nUtterance {msg['utterance_index']} complete: {msg['total_sentences']} sentence(s) generated")
+                return
+            elif msg_type == "error":
+                print(f"  ERROR: {msg['message']}")
+            else:
+                print(f"  Unknown message: {msg}")
+
+
 async def stream_tts(
     url: str,
-    text: str,
+    texts: list[str],
     config: dict,
     output_dir: str,
     simulate_stt: bool = False,
     stt_delay: float = 0.1,
     save_chunks: bool = False,
 ) -> None:
-    """Connect to the streaming TTS endpoint and process audio responses."""
+    """Connect to the streaming TTS endpoint and process audio responses.
+
+    Every text is synthesized over the same connection: input.done flushes an
+    utterance without closing, so only the first one pays the handshake.
+    """
     os.makedirs(output_dir, exist_ok=True)
+    response_format = config.get("response_format", "wav")
+    word_timestamps = bool(config.get("word_timestamps", False))
 
     async with websockets.connect(url) as ws:
-        # 1. Send session config
+        # 1. Send session config; it stays in effect for every utterance below.
         config_msg = {"type": "session.config", **config}
         await ws.send(json.dumps(config_msg))
         print(f"Sent session config: {config}")
 
-        # 2. Send text (either all at once or word-by-word)
-        async def send_text():
-            if simulate_stt:
-                words = text.split(" ")
-                for i, word in enumerate(words):
-                    chunk = word + (" " if i < len(words) - 1 else "")
-                    await ws.send(
-                        json.dumps(
-                            {
-                                "type": "input.text",
-                                "text": chunk,
-                            }
-                        )
-                    )
-                    print(f"  Sent: {chunk!r}")
-                    await asyncio.sleep(stt_delay)
-            else:
-                await ws.send(
-                    json.dumps(
-                        {
-                            "type": "input.text",
-                            "text": text,
-                        }
-                    )
-                )
-                print(f"Sent full text: {text!r}")
-
-            # 3. Signal end of input
-            await ws.send(json.dumps({"type": "input.done"}))
-            print("Sent input.done")
-
-        # Run sender and receiver concurrently
-        sender_task = asyncio.create_task(send_text())
-
-        response_format = config.get("response_format", "wav")
-        word_timestamps = bool(config.get("word_timestamps", False))
-        current_sentence_index = 0
-        current_sentence_text = ""
-        current_chunks: list[bytes] = []
-        current_timestamps: list[dict] = []
-        # Distinguish `null` (aligner failed) from `[]` (silence / no tokens):
-        # the sentence is a failure only if no frame ever carried timestamps.
-        alignment_frame_seen = False
-
-        try:
-            while True:
-                message = await ws.recv()
-
-                if isinstance(message, bytes):
-                    current_chunks.append(message)
-                    print(f"  Received audio chunk for sentence {current_sentence_index}: {len(message)} bytes")
-                else:
-                    # JSON frame
-                    msg = json.loads(message)
-                    msg_type = msg.get("type")
-
-                    if msg_type == "audio.start":
-                        current_sentence_index = msg["sentence_index"]
-                        current_sentence_text = msg.get("sentence_text", "")
-                        current_chunks = []
-                        current_timestamps = []
-                        alignment_frame_seen = False
-                        print(f"  [sentence {msg['sentence_index']}] Generating: {current_sentence_text!r}")
-                    elif msg_type == "audio.chunk":
-                        audio = base64.b64decode(msg["audio_b64"])
-                        current_chunks.append(audio)
-                        chunk_timestamps = msg.get("timestamps")
-                        # Sentence-level alignment sends a trailing timestamp-only
-                        # frame with empty audio; don't write an empty chunk file.
-                        if save_chunks and audio:
-                            chunk_base = os.path.join(
-                                output_dir,
-                                f"sentence_{msg['sentence_index']:03d}_chunk_{msg['chunk_id']:03d}",
-                            )
-                            chunk_pcm = f"{chunk_base}.pcm"
-                            chunk_wav = f"{chunk_base}.wav"
-                            chunk_json = f"{chunk_base}_timestamps.json"
-                            with open(chunk_pcm, "wb") as f:
-                                f.write(audio)
-                            save_pcm_wav(chunk_wav, audio, int(msg.get("sample_rate", 24000)))
-                            with open(chunk_json, "w", encoding="utf-8") as f:
-                                json.dump(
-                                    {
-                                        "sentence_index": msg["sentence_index"],
-                                        "chunk_id": msg["chunk_id"],
-                                        "chunk_start_ms": msg.get("chunk_start_ms"),
-                                        "chunk_end_ms": msg.get("chunk_end_ms"),
-                                        "sentence_text": current_sentence_text,
-                                        "covered_text": timestamp_words_text(chunk_timestamps),
-                                        "timestamps": chunk_timestamps,
-                                    },
-                                    f,
-                                    ensure_ascii=False,
-                                    indent=2,
-                                )
-                        if chunk_timestamps is None:
-                            print(
-                                f"  [sentence {msg['sentence_index']}] chunk {msg['chunk_id']}: "
-                                f"{len(audio)} bytes, timestamps unavailable"
-                            )
-                        else:
-                            alignment_frame_seen = True
-                            current_timestamps.extend(chunk_timestamps)
-                            print(
-                                f"  [sentence {msg['sentence_index']}] chunk {msg['chunk_id']}: "
-                                f"{len(audio)} bytes, {len(chunk_timestamps)} timestamp(s)"
-                            )
-                        if save_chunks and audio:
-                            print(f"    saved chunk -> {chunk_wav}")
-                    elif msg_type == "audio.done":
-                        filename = save_audio_file(
-                            output_dir,
-                            msg["sentence_index"],
-                            response_format,
-                            current_chunks,
-                        )
-                        if word_timestamps:
-                            # `null` on failure, else the accumulated list (`[]` for silence).
-                            sentence_timestamps = current_timestamps if alignment_frame_seen else None
-                            ts_filename = os.path.join(
-                                output_dir,
-                                f"sentence_{msg['sentence_index']:03d}_timestamps.json",
-                            )
-                            with open(ts_filename, "w", encoding="utf-8") as f:
-                                json.dump(
-                                    {
-                                        "sentence_index": msg["sentence_index"],
-                                        "sentence_text": current_sentence_text,
-                                        "covered_text": timestamp_words_text(sentence_timestamps),
-                                        "timestamps": sentence_timestamps,
-                                    },
-                                    f,
-                                    ensure_ascii=False,
-                                    indent=2,
-                                )
-                        print(
-                            f"  [sentence {msg['sentence_index']}] Done"
-                            f" bytes={msg.get('total_bytes', len(b''.join(current_chunks)))}"
-                            f" error={msg.get('error', False)}"
-                            f" -> {filename}"
-                        )
-                        if word_timestamps:
-                            print(f"  [sentence {msg['sentence_index']}] Timestamps -> {ts_filename}")
-                        current_chunks = []
-                        current_timestamps = []
-                        alignment_frame_seen = False
-                    elif msg_type == "session.done":
-                        print(f"\nSession complete: {msg['total_sentences']} sentence(s) generated")
-                        break
-                    elif msg_type == "error":
-                        print(f"  ERROR: {msg['message']}")
-                    else:
-                        print(f"  Unknown message: {msg}")
-        finally:
-            sender_task.cancel()
+        for text in texts:
+            # 2. Send text (either all at once or word-by-word), then flush.
+            sender_task = asyncio.create_task(send_utterance(ws, text, simulate_stt, stt_delay))
             try:
-                await sender_task
-            except asyncio.CancelledError:
-                pass  # Task cancellation is expected during shutdown
+                await receive_utterance(
+                    ws,
+                    output_dir=output_dir,
+                    response_format=response_format,
+                    word_timestamps=word_timestamps,
+                    save_chunks=save_chunks,
+                )
+            finally:
+                sender_task.cancel()
+                try:
+                    await sender_task
+                except asyncio.CancelledError:
+                    pass  # Task cancellation is expected during shutdown
+
+        # 3. Release the connection once there is nothing left to say.
+        await ws.send(json.dumps({"type": "session.close"}))
+        print("Sent session.close")
 
     print(f"\nAudio files saved to: {output_dir}/")
 
@@ -286,7 +325,8 @@ def main():
     parser.add_argument(
         "--text",
         required=True,
-        help="Text to synthesize",
+        action="append",
+        help="Text to synthesize; repeat to send several utterances over one connection",
     )
     parser.add_argument(
         "--output-dir",
@@ -385,7 +425,7 @@ def main():
     asyncio.run(
         stream_tts(
             url=args.url,
-            text=args.text,
+            texts=args.text,
             config=config,
             output_dir=args.output_dir,
             simulate_stt=args.simulate_stt,

pyproject.toml

@@ -75,6 +75,7 @@ dev = [
 ]
 
 demo = [
+    "gradio>=6.7.0",
     "opencv-python>=4.12.0.88",  # cli/stream_client frame reading
     "requests>=2.28.0",  # cli/stream_client HTTP
     "cn2an>=0.5.22",  # Chinese number normalization for TTS (e.g. "1次" -> "一次")
@@ -95,6 +96,12 @@ quack = [
     "quack-kernels>=0.3.11",
 ]
 
+# Optional Blackwell-native FlashAttention-4 backend. Keep this opt-in because
+# the CuTe/CUTLASS dependency is CUDA 13 and Blackwell specific.
+fa4 = [
+    "flash-attn-4[cu13]==4.0.0b18",
+]
+
 docs = [
     "mkdocs>=1.5.0",
     "mkdocs-api-autonav",

realtime_video_prompt_interaction_protocol.md

@@ -1,138 +0,0 @@
-# Realtime Video Prompt Interaction Protocol
-
-This document specifies the prompt-update subset of the realtime video
-interaction protocol exposed by:
-
-```text
-WebSocket /v1/realtime/video
-```
-
-A WebSocket connection contains exactly one video-generation request. Clients
-therefore do not send a request or session ID with an interaction. The server
-assigns a `request_id` in `video.start` and includes it in subsequent
-server-to-client messages.
-
-## Interaction Request
-
-After `video.start`, a client may submit a prompt update:
-
-```json
-{
-  "type": "session.interaction",
-  "interaction": {
-    "event_id": "ui-update-42",
-    "event": {
-      "prompt": "A snowy street illuminated by warm lanterns"
-    },
-    "transition_chunks": 3
-  }
-}
-```
-
-`interaction` has the following fields:
-
-- `event_id` (optional string): session-unique idempotency key. The
-  server preserves a client-provided value or generates a UUID when omitted.
-- `event.prompt` (required string): non-empty target prompt.
-  over which the prompt reaches its target.
-- `transition_chunks` (optional non-negative integer): model-generation chunks
-  over which the prompt reaches its target. If omitted, the model's default transition is used.
-
-Note: Although `["interaction"]["event"]` only contains `prompt`, still consider using `OmniTextPrompt` for this part (with total=false) so that other fields will be supported in the future. Currently, `vllm_omni/diffusion/worker/diffusion_model_runner.py` already takes care of enforcing subset format, routing to meaningful input handing routines, and raising `NotImplementedError`. Only need to adapt this design to the input data structure above.
-
-### Transition Semantics
-
-Prompt transitions are target updates and begin at the next eligible model
-generation boundary after the interaction is processed.
-
-- `transition_chunks = 0` is a hard update: the first affected generation
-  chunk uses the target prompt.
-- `transition_chunks = N`, where `N > 0`, reaches the target in the `N`th
-  affected generation chunk. For example, a three-chunk linear transition uses
-  progress values `1/3`, `2/3`, and `1`.
-
-### Idempotency and Replacement
-
-Within a session:
-
-- repeating an `event_id` with the same payload is a retry and must not
-  apply the update twice;
-- repeating an `event_id` with a different payload is rejected with
-  `event_id_conflict`;
-- a newer prompt target supersedes any older pending prompt target;
-- when a new target interrupts an active transition, its transition starts
-  from the prompt state resolved at its effective boundary.
-
-## Queue Acknowledgement
-
-The server quickly acknowledges admission to the interaction queue:
-
-```json
-{
-  "type": "session.interaction.queued",
-  "request_id": "video_stream-...",
-  "event_id": "ui-update-42"
-}
-```
-
-`queued` confirms syntax validation, idempotency handling, and queue admission.
-It does not claim that the model has scheduled or applied the update. A later
-failure is reported as a correlated error:
-
-```json
-{
-  "type": "error",
-  "request_id": "video_stream-...",
-  "event_id": "ui-update-42",
-  "code": "interaction_failed",
-  "message": "The active model does not support prompt updates"
-}
-```
-
-## Video Chunk Metadata
-
-Each emitted binary payload is preceded by a JSON metadata message. The
-metadata message and its binary payload form one ordered pair and must not be
-interleaved with another outbound message.
-
-```json
-{
-  "type": "video.chunk_metadata",
-  "request_id": "video_stream-...",
-  "kind": "media",
-  "transport_chunk_index": 7,
-  "generation_chunk_index": 6,
-  "num_frames": 9,
-  "byte_length": 184320,
-  "started_event_ids": ["ui-update-42"],
-  "active_event_ids": ["ui-update-42"],
-  "completed_event_ids": []
-}
-```
-
-- `transport_chunk_index` identifies the binary WebSocket payload.
-- `generation_chunk_index` identifies the model-generation chunk represented
-  by the payload.
-- `started_event_ids` lists interactions that first affect this
-  generation chunk.
-- `active_event_ids` lists interactions contributing to this generation
-  chunk, including an ongoing transition.
-- `completed_event_ids` lists interactions that reach their target in
-  this generation chunk.
-
-Encoder trailer bytes are emitted as:
-
-```json
-{
-  "type": "video.chunk_metadata",
-  "request_id": "video_stream-...",
-  "kind": "trailer",
-  "transport_chunk_index": 8,
-  "generation_chunk_index": null,
-  "num_frames": 0,
-  "byte_length": 824,
-  "started_event_ids": [],
-  "active_event_ids": [],
-  "completed_event_ids": []
-}
-```

recipes/JD/JoyAI-VL-Interaction.md

@@ -168,7 +168,7 @@ the model + orchestrator + WebUI together — point `WEBUI_DIR` at your clone.
 python examples/online_serving/joyvl_interaction/cli/run_cli_demo.py \
   path/to/video.mp4 --query "Alert me if a fire breaks out"
 
-pytest tests/fullduplex   # framework + JoyVL unit tests
+pytest tests/e2e/features/fullduplex   # framework + JoyVL unit tests
 ```
 
 ## Testing with an RTSP stream (optional)

recipes/LTX/LTX-2.md

@@ -165,6 +165,40 @@ Pass these fields in online `extra_params` or offline `--extra-body`:
 --extra-body '{"video_cfg_scale":3.0,"audio_cfg_scale":7.0}'
 ```
 
+### Guidance Parallelism
+
+For LTX, `cfg_parallel_size` is the number of ranks used to execute the
+complete guidance plan in parallel. Despite the legacy `cfg` name, it covers
+text CFG, STG, and cross-modality guidance passes; guidance rescaling is
+applied after all pass predictions are gathered.
+
+The default one-stage recipe has four Transformer passes per denoise step:
+`cond`, `uncond`, `ptb` (STG), and `mod` (cross-modality). The useful balanced
+configurations are therefore:
+
+| `--cfg-parallel-size` | Passes per rank | Guidance-slot utilization | Notes |
+|---:|---:|---:|---|
+| `1` | 4 | 100% | Single-rank fused guidance batch |
+| `2` | 2 | 100% | Recommended two-rank configuration |
+| `4` | 1 | 100% | One guidance pass per rank |
+
+Other positive sizes are accepted. When the pass count is not divisible by
+`cfg_parallel_size`, ranks are padded to an equal number of execution slots
+and LTX emits a warning with the expected utilization. For example, four
+passes on three ranks use six slots and have 66.7% guidance-slot utilization.
+
+Start an LTX-2.3 server with two-way guidance parallelism:
+
+```bash
+vllm serve diffusers/LTX-2.3-Diffusers --omni \
+  --cfg-parallel-size 2 --stage-init-timeout 600
+```
+
+The total device count is the product of `cfg_parallel_size` and the other
+configured parallel dimensions. Distilled pipelines use positive-only
+guidance, so `cfg_parallel_size > 1` does not provide guidance speedup for
+them.
+
 ### Python API
 
 Normal per-request values belong in `OmniDiffusionSamplingParams`. Pass the
@@ -273,8 +307,8 @@ bundled offline CLI do not currently expose `sigmas`.
 - Fused one-stage requests must resolve to identical LTX guidance and sigma
   schedules. Keep concurrent requests guidance-homogeneous or use
   `--max-num-seqs 1`.
-- `--cfg-parallel-size 2` supports only the CFG-only plan: disable STG and
-  modality guidance and set both rescale fields to `0.0`.
+- `--cfg-parallel-size` shards the complete LTX guidance plan, including STG,
+  modality guidance, and rescale-compatible prediction gathering.
 
 ## Operational Notes
 

recipes/MiniMaxAI/MiniMax-H3-NPU.md

@@ -0,0 +1,162 @@
+# MiniMax H3 on Ascend NPU
+
+> Joint video and audio generation with text, first-frame, image/audio, or
+> multi-video conditions — Ascend NPU deployment guide
+
+## Summary
+
+- Vendor: MiniMaxAI
+- Model: [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3)
+- Tasks: T2VA, FL2VA, and Ref2VA
+- Mode: OpenAI-compatible `/v1/videos` HTTP serving
+- Hardware: Atlas 800I A3
+- Maintainer: Community
+
+This recipe adapts [MiniMax-H3.md](MiniMax-H3.md) for Ascend NPU
+environments. Differences from the GPU path:
+
+- Runs on a CPU-only (aarch64) PyTorch build with `torch_npu`; no CUDA
+  runtime is present.
+- Audio loading does **not** require TorchCodec (whose aarch64 wheels are
+  built against CUDA torch and fail to load on CPU-only builds). vLLM-Omni
+  automatically falls back to soundfile / ffmpeg for wav/mp3/m4a/mp4 audio
+  inputs.
+
+## Prerequisites
+
+### Checkpoint
+
+Same as the GPU recipe — Hugging Face access approval is required:
+
+```bash
+hf auth login
+export MODEL_ROOT=/path/to/MiniMax-H3
+hf download MiniMaxAI/MiniMax-H3 --local-dir "${MODEL_ROOT}"
+```
+
+### Environment
+
+- Ascend driver & firmware: 25.5.1
+- CANN toolkit: 9.0.1
+- Python: 3.12
+- PyTorch: 2.10.0+cpu
+- torch_npu: 2.10.0.post2
+- Install vLLM-Omni from a checkout with MiniMax H3 support:
+
+```bash
+uv venv
+source .venv/bin/activate
+uv pip install -e .
+```
+
+Install the **mindie-sd** operator library to enable Ascend-optimized fused
+operators (`adalayernorm`, etc.):
+
+```bash
+git clone https://gitcode.com/Ascend/MindIE-SD.git && cd MindIE-SD
+
+# Comment out the tik_ops build step (not needed for this use case)
+sed -i 's|^\(\s*\)source ${current_script_dir}/build_tik_ops.sh|\1# source ${current_script_dir}/build_tik_ops.sh|' build/build_ops.sh
+
+python setup.py bdist_wheel
+cd dist
+pip install mindiesd-*.whl
+```
+
+- `ffmpeg` and `ffprobe` must be available on `PATH`. They are used for
+  reference-video preparation and MP4 output.
+- Reference-video decoding uses `decord` when available and falls back to
+  PyAV otherwise.
+- Audio inputs do not require TorchCodec: wav/mp3/m4a/mp4 files are loaded
+  at native sample rate through the soundfile / ffmpeg fallback.
+
+## Start a server
+
+One server loads one checkpoint partition. Set `MODEL` to `FL2VA` for T2VA
+and FL2VA requests, or to `Ref2VA` for Ref2VA requests.
+
+### Multi-NPU: 768P validated configuration
+
+Validated on eight NPUs of an Atlas 800I A3 server with Ulysses sequence
+parallelism degree 8, text-encoder tensor parallelism degree 8, native tiled
+VAE patch parallelism degree 8, and layerwise offload:
+
+```bash
+export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+export PORT=9098
+export MODEL="${MODEL_ROOT}/FL2VA"
+export VLLM_WORKER_MULTIPROC_METHOD=spawn
+export VLLM_OMNI_VIDEO_SYNC_TIMEOUT=1800
+export PYTHONDONTWRITEBYTECODE=1
+export HF_HUB_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+vllm serve "${MODEL}" \
+  --omni \
+  --host 0.0.0.0 \
+  --port "${PORT}" \
+  --trust-remote-code \
+  --num-gpus 8 \
+  --usp 8 \
+  --ring 1 \
+  --text-encoder-tp-size 8 \
+  --enable-layerwise-offload \
+  --vae-parallel-mode tile \
+  --vae-use-tiling \
+  --vae-patch-parallel-size 8 \
+  --diffusion-attention-backend FLASH_ATTN
+```
+
+Do not add `--enforce-eager`. The first request includes regional
+compilation; warm the server once before measuring steady-state latency.
+H3 is CFG-distilled, so `--cfg-parallel-size` must remain 1.
+
+To serve Ref2VA, stop the FL2VA server and restart with:
+
+```bash
+export MODEL="${MODEL_ROOT}/Ref2VA"
+```
+
+## HTTP API examples
+
+The request format is identical to the GPU recipe; see
+[MiniMax-H3.md § HTTP API examples](MiniMax-H3.md#http-api-examples).
+Use the validated 768P shapes (e.g. `width=1344 height=768`) on NPU.
+
+## Key parameters
+
+Same as the GPU recipe; see
+[MiniMax-H3.md § Key parameters](MiniMax-H3.md#key-parameters).
+The validated resolution on NPU is 768P (e.g. 1344x768).
+
+## Validated NPU evidence
+
+Measured on an Atlas 800I A3 server (8x NPU) with CANN 9.0.1,
+PyTorch 2.10.0+cpu, and torch_npu 2.10.0.post2, using the multi-NPU
+configuration above:
+
+| Workload | Configuration |
+|----------|---------------|
+| T2VA, 209 frames, 1344x768 | TE TP8, layerwise offload, Ulysses 8, VPP8 tile, regional compile |
+| Ref2VA (prompt + video), 124 frames, 1344x768 | TE TP8, layerwise offload, Ulysses 8, VPP8 tile, regional compile |
+
+These measurements describe the validated shapes rather than a general
+throughput guarantee.
+
+## Known limitations
+
+- Each server process loads only one checkpoint partition.
+- H3 currently executes one generation request per diffusion batch.
+- The first regional-compile request is a warmup and should not be included
+  in steady-state performance measurements.
+- Image+audio Ref2VA accepts exactly one image and one audio reference.
+- Video Ref2VA accepts one or more video files, but not an additional
+  standalone audio reference.
+- VAE patch parallelism requires size 1 or the full DiT group size and
+  supports the H3 native `tile` mode only.
+
+## Additional resources
+
+- [MiniMax-H3.md](MiniMax-H3.md) — full GPU guide
+- [Supported models](../../docs/models/supported_models.md)
+- [Video API](../../docs/serving/videos_api.md)

recipes/MiniMaxAI/MiniMax-H3.md

@@ -0,0 +1,374 @@
+# MiniMax H3
+
+> Joint video and audio generation with text, first-frame, image/audio, or
+> multi-video conditions
+
+## Summary
+
+- Vendor: MiniMaxAI
+- Model: [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3)
+- Tasks: T2VA, FL2VA, and Ref2VA
+- Mode: OpenAI-compatible `/v1/videos` HTTP serving
+- Maintainer: Community
+
+MiniMax H3 is a CFG-distilled joint video/audio diffusion transformer. Its
+checkpoint has two independently served partitions:
+
+- `FL2VA`: text-to-video+audio (`t2va`) and first-frame-to-video+audio
+  (`fl2va`)
+- `Ref2VA`: image+audio or one-or-more-video reference-to-video+audio
+  (`ref2va`)
+
+The generated MP4 contains H.264 video and synchronized stereo audio.
+
+## Prerequisites
+
+The checkpoint requires Hugging Face access approval. Download it and point
+`MODEL_ROOT` at the directory containing the `FL2VA` and `Ref2VA`
+subdirectories:
+
+```bash
+hf auth login
+export MODEL_ROOT=/path/to/MiniMax-H3
+hf download MiniMaxAI/MiniMax-H3 --local-dir "${MODEL_ROOT}"
+```
+
+Install vLLM-Omni from the checkout containing MiniMax H3 support. On
+Blackwell, install the optional FlashAttention-4 dependency:
+
+```bash
+uv venv
+source .venv/bin/activate
+uv pip install -e '.[fa4]'
+```
+
+`ffmpeg` and `ffprobe` must be available on `PATH`. They are used for
+reference-video preparation and MP4 output.
+
+## Start a server
+
+One server loads one checkpoint partition. Set `MODEL` to `FL2VA` for T2VA
+and FL2VA requests, or to `Ref2VA` for either Ref2VA request.
+
+### Single GPU: accuracy and memory first
+
+The single-GPU configuration uses model-level CPU offload.
+This matches the accuracy-qualified reference path and prevents the Qwen3-VL
+encoder and DiT from being resident on the GPU at the same time.
+
+```bash
+export MODEL="${MODEL_ROOT}/FL2VA"
+export PORT=8091
+
+CUDA_VISIBLE_DEVICES=0 \
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_OMNI_VIDEO_SYNC_TIMEOUT=1800 \
+vllm serve "${MODEL}" \
+  --omni \
+  --host 0.0.0.0 \
+  --port "${PORT}" \
+  --trust-remote-code \
+  --num-gpus 1 \
+  --enable-cpu-offload \
+  --diffusion-attention-backend FLASH_ATTN
+```
+
+Use a GPU with enough memory for the active H3 component and enough system RAM
+for the offloaded components. CPU offload reduces GPU memory pressure but adds
+PCIe/NVLink transfer latency.
+
+### Four GPUs: measured best-practice throughput
+
+The best validated four-GPU configuration on four NVIDIA B300 GPUs is:
+
+- no CPU or layerwise offload;
+- Ulysses sequence parallelism degree 4;
+- native tiled VAE patch parallelism degree 4;
+- regional `torch.compile` for the repeated DiT blocks;
+- FlashAttention, with Ring and TP left at 1.
+
+```bash
+export MODEL="${MODEL_ROOT}/FL2VA"
+export PORT=8091
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+FLASHINFER_DISABLE_VERSION_CHECK=1 \
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_OMNI_VIDEO_SYNC_TIMEOUT=1800 \
+vllm serve "${MODEL}" \
+  --omni \
+  --host 0.0.0.0 \
+  --port "${PORT}" \
+  --trust-remote-code \
+  --num-gpus 4 \
+  --usp 4 \
+  --ring 1 \
+  --vae-patch-parallel-size 4 \
+  --vae-parallel-mode tile \
+  --vae-use-tiling \
+  --diffusion-attention-backend FLASH_ATTN
+```
+
+Do not add `--enforce-eager` to this performance configuration. The first
+request includes regional compilation; warm the server once before measuring
+steady-state latency. H3 is CFG-distilled, so `--cfg-parallel-size` must remain
+1. The H3 VAE supports its native `tile` mode, not
+`spatial_shard_height` or `spatial_shard_width`.
+
+### Text encoder tensor parallelism
+
+The Qwen3-VL text encoder (~51.5 GB in BF16 for the retained 50 layers) is by
+default fully resident on the DiT main rank.  On multi-GPU no-offload runs that
+rank becomes the peak-memory hotspot.  Add `--text-encoder-tp-size N` to shard
+the encoder across the first `N` DiT ranks (the encoder is implemented with
+vLLM-style tensor-parallel layers and runs with distributed collectives over
+its own encoder process group):
+
+```bash
+export MODEL="${MODEL_ROOT}/FL2VA"
+export PORT=8091
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+FLASHINFER_DISABLE_VERSION_CHECK=1 \
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_OMNI_VIDEO_SYNC_TIMEOUT=1800 \
+vllm serve "${MODEL}" \
+  --omni \
+  --host 0.0.0.0 \
+  --port "${PORT}" \
+  --trust-remote-code \
+  --num-gpus 4 \
+  --usp 4 \
+  --ring 1 \
+  --text-encoder-tp-size 4 \
+  --vae-patch-parallel-size 4 \
+  --vae-parallel-mode tile \
+  --vae-use-tiling \
+  --diffusion-attention-backend FLASH_ATTN
+```
+
+`N` must divide the Qwen3-VL head counts (64 attention heads / 8 KV heads), so
+valid values on a 4-GPU server are 1, 2, and 4 (1, 2, 4, 8 on 8 GPUs).  The
+encoder TP rank set is the first `N` DiT ranks; on 4 GPUs
+`--text-encoder-tp-size 4` shards the encoder 4-way, dropping the DiT main
+rank's no-offload peak by roughly `(N-1)/N` of the ~51.5 GB encoder while the
+other ranks each gain `~51.5/N` GB.  The encoder output remains identical to
+the reference path within bf16 rounding: every encoder rank all-reduces the
+row-parallel projections, so the full `[seq, 5120]` layer-50 hidden state is
+replicated on every rank.
+
+To serve Ref2VA, stop the FL2VA server and restart the same one- or four-GPU
+command with:
+
+```bash
+export MODEL="${MODEL_ROOT}/Ref2VA"
+```
+
+### Online FP8 quantization
+
+MiniMax H3 supports load-time FP8 quantization of the DiT. The checkpoint
+remains BF16 on disk; vLLM-Omni quantizes eligible weights while loading and
+uses dynamic activation scaling during inference. By default, attention and
+MLP linears in the token refiner and main DiT blocks, the condition
+projection, and all AdaLN projections use FP8. Patch, timestep, and final
+projections remain FP32; the text encoder and VAEs are unchanged.
+
+Add this option to an existing H3 server command:
+
+```bash
+--quantization fp8
+```
+
+Use `ignored_layers` to keep any otherwise eligible linear in BF16. H3
+resolves the `transformer` component before constructing the DiT, so names do
+not start with `transformer.`. Entries are exact runtime linear prefixes; a
+parent name such as `blocks.0.attn` does not exclude its children.
+
+Eligible names are the `attn.qkv_proj`, `attn.out_proj`, `mlp.fc1`, and
+`mlp.fc2` children under `token_refiner.blocks.<0-1>` or `blocks.<0-49>`.
+The other eligible names are `condition_proj`,
+`blocks.<0-49>.adaln_proj.linear`, and `final_layer.adaln_proj.linear`.
+For example, keep the first main block's attention projections in BF16 with:
+
+```bash
+--diffusion-quantization-config \
+  '{"transformer":{"method":"fp8","ignored_layers":["blocks.0.attn.qkv_proj","blocks.0.attn.out_proj"]}}'
+```
+
+The structured option replaces `--quantization fp8`. Online FP8 is currently
+incompatible with H3 layerwise offload because the offload path produces a
+weight stride rejected by the Cutlass FP8 kernel. Use resident FP8 with tensor
+parallelism and VAE tiling instead.
+
+## HTTP API examples
+
+The following requests use the synchronous endpoint so the returned body can
+be saved directly as an MP4. The asynchronous `POST /v1/videos` endpoint can
+also be used when job polling is preferred.
+
+All four tasks use 24 FPS, 50 sigma points, seed values from the validated
+workloads, and the checkpoint-reference video/audio flow shifts of 12 and 3.
+Decimal durations are passed through `extra_params`.
+
+Set the endpoint once:
+
+```bash
+export API_URL="http://127.0.0.1:${PORT}/v1/videos/sync"
+```
+
+### 1. T2VA: text to video and audio
+
+Run this request against the `FL2VA` partition:
+
+```bash
+curl -sS -X POST "${API_URL}" \
+  -F 'prompt=In a snowy blue-purple forest, Ori carefully walks past a sleeping giant; footsteps crunch in the snow while the creature breathes and softly snorts.' \
+  -F 'width=1344' \
+  -F 'height=768' \
+  -F 'fps=24' \
+  -F 'num_inference_steps=50' \
+  -F 'flow_shift=12' \
+  -F 'seed=1101' \
+  -F 'extra_params={"task":"t2va","duration":8.7,"audio_flow_shift":3.0}' \
+  -o t2va.mp4
+```
+
+### 2. FL2VA: first frame to video and audio
+
+Run this request against the `FL2VA` partition. When width and height are
+omitted, H3 preserves the first-frame aspect ratio and uses a 768-pixel short
+edge.
+
+```bash
+export FIRST_FRAME=/path/to/fl2va_first_frame.png
+
+curl -sS -X POST "${API_URL}" \
+  -F 'prompt=A man stands beside a yellow car at night. The car drives away; he follows it with his eyes and begins singing sadly, with synchronized voice and city ambience.' \
+  -F 'fps=24' \
+  -F 'num_inference_steps=50' \
+  -F 'flow_shift=12' \
+  -F 'seed=2101' \
+  -F 'extra_params={"task":"fl2va","duration":8.7,"audio_flow_shift":3.0}' \
+  -F "input_reference=@${FIRST_FRAME};type=image/png" \
+  -o fl2va.mp4
+```
+
+### 3. Ref2VA: image and audio reference
+
+Run this request against the `Ref2VA` partition. `audio_reference` accepts an
+HTTP(S) URL or a `data:` URL. In one terminal, expose the local reference
+assets to the serving host:
+
+```bash
+python -m http.server 8092 \
+  --bind 127.0.0.1 \
+  --directory /path/to/reference_assets
+```
+
+Then submit the image and audio request from another terminal:
+
+```bash
+export REF_IMAGE=/path/to/reference_assets/ref2va_image.png
+export AUDIO_URL=http://127.0.0.1:8092/ref2va_audio.mp3
+
+curl -sS -X POST "${API_URL}" \
+  -F 'prompt=A white cat with black mustache and eyebrow markings sits on a beige couch, lip-syncing precisely to the complete reference audio before shifting from confusion to deadpan speechlessness.' \
+  -F 'width=1344' \
+  -F 'height=768' \
+  -F 'fps=24' \
+  -F 'num_inference_steps=50' \
+  -F 'flow_shift=12' \
+  -F 'seed=3101' \
+  -F 'extra_params={"task":"ref2va","duration":15.0,"audio_flow_shift":3.0}' \
+  -F "input_reference=@${REF_IMAGE};type=image/png" \
+  -F "audio_reference={\"audio_url\":\"${AUDIO_URL}\"}" \
+  -o ref2va_image_audio.mp4
+```
+
+The requested duration should cover the complete audio. If `duration` is
+shorter, the reference soundtrack is truncated to the generated clip.
+
+### 4. Ref2VA: two video references
+
+Run this request against the `Ref2VA` partition. Repeat the
+`input_references` multipart field once per source video. H3 consumes the
+videos in form order and preserves their original soundtracks during
+conditioning.
+
+```bash
+export SUBJECT_VIDEO=/path/to/green_screen_subject.mp4
+export BACKGROUND_VIDEO=/path/to/fairytale_background.mov
+
+curl -sS -X POST "${API_URL}" \
+  -F 'prompt=Remove the green screen background of Video 1 and replace it with the fairytale environment from Video 2. Match the background motion to the character actions and relight the character to fit the scene.' \
+  -F 'width=1344' \
+  -F 'height=768' \
+  -F 'fps=24' \
+  -F 'num_inference_steps=50' \
+  -F 'flow_shift=12' \
+  -F 'seed=3101' \
+  -F 'extra_params={"task":"ref2va","duration":15.0,"audio_flow_shift":3.0}' \
+  -F "input_references=@${SUBJECT_VIDEO};type=video/mp4" \
+  -F "input_references=@${BACKGROUND_VIDEO};type=video/quicktime" \
+  -o ref2va_video_video.mp4
+```
+
+The server stores uploaded multi-video references only for the lifetime of the
+request and deletes the temporary files after generation. Video Ref2VA uses
+the source-video soundtracks and does not accept a separate
+`audio_reference`.
+
+## Key parameters
+
+| Parameter | Recommended value | Notes |
+|-----------|-------------------|-------|
+| `task` | `t2va`, `fl2va`, or `ref2va` | Passed in `extra_params`; must match the served partition |
+| `duration` | Workload-specific | Decimal seconds in `extra_params`; converted to H3-compatible frame count |
+| `fps` | `24` | H3 output FPS is fixed |
+| `num_inference_steps` | `50` | Matches the reference accuracy workloads |
+| `flow_shift` | `12` | Video sigma shift |
+| `audio_flow_shift` | `3` | Audio sigma shift, passed in `extra_params` |
+| `seed` | Task-specific | Use a fixed value for reproducibility |
+| `width`, `height` | Multiples of 32 | Aspect ratio must be between 1:4 and 4:1 |
+
+## Validated four-GPU evidence
+
+The four-GPU recommendation was measured on four NVIDIA B300 GPUs with one
+excluded warmup followed by three requests.
+
+| Workload | Configuration | Observed result |
+|----------|---------------|-----------------|
+| FL2VA, 209 frames, 1248x768 | no offload, U4, VPP4 tile, regional compile | 86.964 s mean HTTP client latency |
+| Two-video Ref2VA, 362 frames, 1344x768 | no offload, U4, VPP4 tile, regional compile | 784.394 s accounted model-stage mean |
+
+These measurements describe the validated shapes rather than a general
+throughput guarantee. Multi-video Ref2VA is much slower because the two
+reference videos expand both the Qwen3-VL vision sequence and the packed DiT
+attention sequence.
+
+## Validated FP8 evidence
+
+With eager DiT/text-encoder TP2 and VAE tiling, the 384x672, 107-frame,
+10-step quality case measured LPIPS 0.1156 (limit 0.20), PSNR 23.6316 dB,
+audio spectral cosine 0.9589 (minimum 0.80), and audio RMS ratio 0.9342. The
+resident per-GPU peak was 68.52 GiB for BF16 and 53.51 GiB for FP8, a 22%
+reduction.
+
+## Known limitations
+
+- Each server process loads only one checkpoint partition.
+- H3 currently executes one generation request per diffusion batch.
+- The first regional-compile request is a warmup and should not be included in
+  steady-state performance measurements.
+- Online FP8 is not compatible with layerwise offload.
+- Image+audio Ref2VA accepts exactly one image and one audio reference.
+- Video Ref2VA accepts one or more video files, but not an additional standalone
+  audio reference.
+- VAE patch parallelism requires size 1 or the full DiT group size and supports
+  the H3 native `tile` mode only.
+
+## Additional resources
+
+- [Supported models](../../docs/models/supported_models.md)
+- [Video API](../../docs/serving/videos_api.md)
+- [Diffusion parallelism](../../docs/user_guide/diffusion/parallelism/overview.md)

recipes/NVIDIA/GR00T-N1.7.md

@@ -69,8 +69,13 @@ Notes:
 ## Verification
 
 ```python
-from tests.gr00t.openpi_client_helper import run_policy_session, validate_session_result
-validate_session_result(run_policy_session(host="127.0.0.1", port=8000))
+from tests.helpers.runtime import OpenAIClientHandler
+
+handler = OpenAIClientHandler(host="127.0.0.1", port=8000, log_stats=False)
+response = handler.send_robot_openpi_ws_request(
+    {"run_default_policy_session": True, "session_id": "gr00t-smoke"}
+)[0]
+assert response.operation_responses[-1]["status"] == "reset successful"
 ```
 
 Or run the e2e test suite:

recipes/OpenBMB/MiniCPM-o-4_5.md

@@ -30,8 +30,6 @@ also provided.
   `hf_config.version="4.5"`):
   - Default single-GPU compatibility layout (auto-loaded):
     [`vllm_omni/deploy/minicpmo_4_5.yaml`](../../vllm_omni/deploy/minicpmo_4_5.yaml)
-  - Recommended 2-GPU continuous-batching layout:
-    [`vllm_omni/deploy/minicpmo_4_5_batching.yaml`](../../vllm_omni/deploy/minicpmo_4_5_batching.yaml)
   - 2-GPU and 3-GPU layouts:
     [`vllm_omni/deploy/minicpmo_4_5_2gpu.yaml`](../../vllm_omni/deploy/minicpmo_4_5_2gpu.yaml),
     [`vllm_omni/deploy/minicpmo_4_5_3gpu.yaml`](../../vllm_omni/deploy/minicpmo_4_5_3gpu.yaml)
@@ -68,7 +66,7 @@ Code2Wav consumes them through a shared-memory async connector.
 MiniCPM-o 4.5 now requires the three-stage topology: the Talker owns
 request-local codec generation and Code2Wav owns waveform state and
 reference-voice prompt features. `minicpmo_4_5.yaml` remains the stable
-single-GPU entry point; `minicpmo_4_5_batching.yaml` is the recommended
+single-GPU entry point; `minicpmo_4_5_2gpu.yaml` is the recommended
 two-GPU profile. The removed fused two-stage implementation is not retained as
 a fallback because it would duplicate state machines and correctness paths.
 
@@ -79,7 +77,9 @@ a fallback because it would duplicate state machines and correctness paths.
 The default
 [`vllm_omni/deploy/minicpmo_4_5.yaml`](../../vllm_omni/deploy/minicpmo_4_5.yaml)
 co-locates Thinker, codec-only Talker, and Code2Wav on GPU 0. Their
-`gpu_memory_utilization` budgets are 0.55, 0.22, and 0.22. All stages admit
+`gpu_memory_utilization` budgets are 0.55, 0.15, and 0.15. The remaining
+device memory is left available for runtime workspaces such as the HiFi-GAN
+vocoder's cuDNN kernels. All stages admit
 up to four sequences. Startup video profiling is bounded to 32 frames per
 video. Use the two-GPU profile for production throughput.
 
@@ -105,7 +105,7 @@ The deploy config is auto-loaded by the model registry — no
 For the recommended two-GPU layout, add:
 
 ```bash
---deploy-config vllm_omni/deploy/minicpmo_4_5_batching.yaml
+--deploy-config vllm_omni/deploy/minicpmo_4_5_2gpu.yaml
 ```
 
 #### Performance comparison
@@ -175,16 +175,18 @@ in another choice's `message.audio.data` (24 kHz mono, see Notes). With
 `modalities: ["text", "audio"]` you typically get two `choices` entries
 (one text, one audio).
 
-**Streaming text + speech**:
+**Streaming text + speech** (use `--stream`):
 
 ```bash
-python examples/online_serving/minicpmo/streaming_chat_completion.py \
-    --base-url http://localhost:8099/v1 \
-    --output minicpmo_stream.wav
+python examples/online_serving/minicpmo/openai_chat_completion_client_for_multimodal_generation.py \
+    --query-type text \
+    --prompt "Say hello, then introduce vLLM in one sentence." \
+    --port 8099 \
+    --stream
 ```
 
-The client prints text deltas as they arrive and reconstructs one valid WAV
-from the independently encoded audio deltas.
+The client prints text deltas as they arrive and saves streamed audio chunks
+to WAV files.
 
 **Gradio demo (text + image + audio + video UI)**:
 
@@ -202,9 +204,10 @@ speech output (TTS)"** checkbox on / off.
 
 #### Notes
 
-- Memory budget: Thinker, Talker, and Code2Wav reserve 0.55, 0.22, and
-  0.22 of GPU 0. The larger Thinker share protects its multimodal KV cache;
-  all three model processes still share one CUDA device.
+- Memory budget: Thinker, Talker, and Code2Wav reserve 0.55, 0.15, and
+  0.15 of GPU 0. The larger Thinker share protects its multimodal KV cache,
+  while the unreserved memory remains available to runtime kernels; all three
+  model processes still share one CUDA device.
 - `--trust-remote-code` is required — the HF repo ships a custom
   `MiniCPMO` config / model class.
 - Stage 0 Thinker and Stage 1 Talker enable vLLM CUDA Graphs. Stage 2 remains

recipes/README.md

@@ -45,6 +45,7 @@ recipes/
 | [`inclusionAI/Ming-flash-omni-2.0.md`](./inclusionAI/Ming-flash-omni-2.0.md) | Online serving for multimodal chat + standalone TTS | 4x H100 / 1x H100 80GB |
 | [`inclusionAI/Ming-omni-tts.md`](./inclusionAI/Ming-omni-tts.md) | Offline + online dense Ming TTS/audio generation | 1x H100 80GB / 1x AMD MI300X (ROCm 7.2) |
 | [`IndexTeam/IndexTTS-2.md`](./IndexTeam/IndexTTS-2.md) | Online serving for voice-cloned TTS with optional emotion control | 1x L4 24GB or larger CUDA GPU |
+| [`MiniMaxAI/MiniMax-H3.md`](./MiniMaxAI/MiniMax-H3.md) | T2VA, FL2VA, image+audio Ref2VA, and multi-video Ref2VA serving | 1x GPU with CPU offload / 4x B300 |
 | [`krea/Krea-2.md`](./krea/Krea-2.md) | Text-to-image (Turbo + Raw), offline + online, with LoRA | 1x H100 80GB |
 | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | LTX-2/LTX-2.3 text-to-video and image-to-video with synchronized audio | H200 141GB / 96GB-class GPU |
 | [`MammothModa2/MammothModa2.md`](./MammothModa2/MammothModa2.md) | Preview and Dev text-to-image (AR → DiT); Dev text/image understanding | Preview: 1x L40S 48GB / 1x ≥40GB GPU; Dev: 1x NVIDIA GPU with sufficient cache headroom |

recipes/cosmos3/Cosmos3-DistOffload.md

@@ -0,0 +1,191 @@
+# Cosmos3 — Distributed Layerwise Offload
+
+> Sharding weights across DP ranks with H2D + AllGather overlap for multi-NPU/GPU deployment
+
+## When to use this recipe
+
+Use this recipe when deploying Cosmos3-Super (64B / 124 GB) or Cosmos3-Nano (17B / 33 GB) across multiple devices where the model does not fit on a single device's HBM, or when you need higher throughput via DP multi-concurrency.
+
+**Key benefits:**
+- Each rank stores only 1/dp_size of the model weights on host memory
+- Only 2 transformer blocks reside on each device at any time (fixed double-buffer)
+- DP multi-concurrency: N concurrent requests with near-linear throughput scaling
+- mmap weight loading: shared page cache eliminates O(dp_size × model_size) RSS
+
+## Prerequisites
+
+- 2+ NPU or GPU devices (Ascend 910B3, NVIDIA B300, etc.)
+- vLLM-Omni with distributed layerwise offload support
+- Cosmos3-Nano or Cosmos3-Super checkpoint (local path or HuggingFace repo ID)
+
+## Serve commands
+
+### Cosmos3-Nano with DP=4 + AllGather (recommended for max throughput)
+
+```bash
+vllm serve /path/to/Cosmos3-Nano \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --no-guardrails --init-timeout 3600 \
+  --vae-use-tiling --vae-patch-parallel-size 1 \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4
+```
+
+### Cosmos3-Super with DP=2 + AllGather (for models > single-card HBM)
+
+```bash
+vllm serve /path/to/Cosmos3-Super \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --no-guardrails --init-timeout 3600 \
+  --vae-use-tiling --vae-patch-parallel-size 1 \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 2
+```
+
+### Cosmos3-Super with DP=4 + AllGather
+
+```bash
+vllm serve /path/to/Cosmos3-Super \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --no-guardrails --init-timeout 3600 \
+  --vae-use-tiling --vae-patch-parallel-size 1 \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4
+```
+
+### Without AllGather (full weights per rank, no sharding)
+
+Use `--dlo-no-use-allgather` when you want each rank to load full weights independently —
+no AllGather synchronization, no concurrent request requirement, but N× host memory.
+
+```bash
+vllm serve /path/to/Cosmos3-Nano \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --no-guardrails --init-timeout 3600 \
+  --vae-use-tiling --vae-patch-parallel-size 1 \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4 \
+  --dlo-no-use-allgather
+```
+
+### With SP (sequence parallel) instead of DP
+
+For long-sequence workloads where SP parallelism is preferred:
+
+```bash
+vllm serve /path/to/Cosmos3-Super \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --no-guardrails --init-timeout 3600 \
+  --vae-use-tiling --vae-patch-parallel-size 1 \
+  --enable-distributed-layerwise-offload \
+  --usp 4
+```
+
+## Request examples
+
+### Text-to-Image (T2I)
+
+```bash
+curl -s -o output.png -X POST "http://localhost:8000/v1/images/generations" \
+  -H "Content-Type: multipart/form-data" \
+  -F "model=/path/to/Cosmos3-Nano" \
+  -F "prompt=A robot arm cleaning a plate, cinematic" \
+  -F "size=1024x1024" \
+  -F "num_inference_steps=50" \
+  -F "guidance_scale=6.0" \
+  -F "seed=42"
+```
+
+### Text-to-Video (T2V)
+
+```bash
+curl -s -o output.mp4 -X POST "http://localhost:8000/v1/videos/sync" \
+  -H "Accept: video/mp4" \
+  -F "model=/path/to/Cosmos3-Nano" \
+  -F "prompt=A robot arm cleaning a plate, cinematic shot" \
+  -F "negative_prompt=blurry" \
+  -F "size=832x480" \
+  -F "num_frames=29" \
+  -F "fps=24" \
+  -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" \
+  -F "max_sequence_length=4096" \
+  -F "flow_shift=10.0" \
+  -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":false}' \
+  -F "seed=17"
+```
+
+> **Important:** When using DP multi-concurrency (AllGather mode), `num_inference_steps`
+> must be specified explicitly and must be the same for all concurrent requests.
+> `num_inference_steps=None` is not allowed because it may resolve differently
+> per request mode (e.g., T2V=35 vs action_mode=30), which would deadlock AllGather.
+
+### Multiple concurrent requests (DP multi-concurrency)
+
+With `--data-parallel-size 4`, send 4 concurrent requests to maximize throughput:
+
+```bash
+for i in 1 2 3 4; do
+  curl -s -o output_${i}.mp4 -X POST "http://localhost:8000/v1/videos/sync" \
+    -H "Accept: video/mp4" \
+    -F "model=/path/to/Cosmos3-Nano" \
+    -F "prompt=A robot arm cleaning a plate, shot ${i}" \
+    -F "negative_prompt=blurry" \
+    -F "size=832x480" \
+    -F "num_frames=29" \
+    -F "fps=24" \
+    -F "num_inference_steps=35" \
+    -F "guidance_scale=6.0" \
+    -F "max_sequence_length=4096" \
+    -F "flow_shift=10.0" \
+    -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":false}' \
+    -F "seed=17" &
+done
+wait
+```
+
+## CLI flags
+
+| Flag | Description | Default |
+|------|-------------|---------|
+| `--enable-distributed-layerwise-offload` | Enable DLO with H2D + AllGather overlap | `false` |
+| `--data-parallel-size N` | Number of DP ranks (weight sharding + concurrent requests) | `1` |
+| `--dlo-use-allgather` | Use shard + AllGather for weight reconstruction (recommended) | `true` |
+| `--dlo-no-use-allgather` | Each rank loads full weights independently (no sharding, no AllGather) | `false` |
+| `--usp N` | Use SP instead of DP for weight sharding (long sequences) | `1` |
+
+## Mode comparison
+
+| Mode | CLI flags | CPU/rank | HBM/card | Throughput | Use case |
+|------|----------|---------|---------|-----------|----------|
+| DLO + AllGather (DP4) | `--enable-distributed-layerwise-offload --data-parallel-size 4` | 1/4 model | 2 blocks | **3.3× HSDP** | Max throughput, short sequences |
+| DLO + AllGather (DP2) | `--enable-distributed-layerwise-offload --data-parallel-size 2` | 1/2 model | 2 blocks | 2× HSDP | Balanced throughput + memory |
+| DLO + AllGather (SP4) | `--enable-distributed-layerwise-offload --usp 4` | 1/4 model | 2 blocks | 1× (single req) | Long sequences (720p+) |
+| DLO no-AllGather | `--enable-distributed-layerwise-offload --dlo-no-use-allgather` | Full model | 2 blocks | 0.4× HSDP | No AllGather overhead, high CPU |
+| HSDP | `--use-hsdp --hsdp-shard-size 4` | 0 | 1/4 model | 1× (baseline) | Weights fit in HBM |
+
+## Memory expectations
+
+| Model | Config | Host RAM (cgroup) | HBM/card |
+|-------|--------|-------------------|----------|
+| Nano (33 GB) | DP2 + AllGather | ~38 GB | ~10 GB |
+| Nano (33 GB) | DP4 + AllGather | ~47 GB | ~10 GB |
+| Super (124 GB) | DP2 + AllGather | ~157 GB | ~15 GB |
+| Super (124 GB) | DP4 + AllGather | ~172 GB | ~10 GB |
+
+> Host RAM is dominated by page cache (1× model_size, shared across ranks).
+> HBM only holds 2 transformer blocks + framework + activations, independent of model size.
+
+## Limitations
+
+- `--dlo-use-allgather` (default) requires all concurrent requests to have the same
+  `num_inference_steps` (AllGather is a collective that needs all ranks synchronized)
+- Online quantization (FP8) is incompatible with mmap loading — falls back to regular
+  `load_weights()` automatically
+- Tensor Parallel is not supported (DLO uses DP-based sharding, a different dimension)
+- HSDP + AllGather is rejected (would double-shard weights)

recipes/cosmos3/Cosmos3-Edge.md

@@ -17,7 +17,9 @@ The small Cosmos3 checkpoint, for constrained deployments. It runs on the shared
 
 - **Transformer**: `Cosmos3EdgeVFMTransformer` (Nemotron-based), auto-selected from
   `backbone_type: cosmos3_edge_nemotron_dense` — nothing to pass.
-- **Defaults** differ; omit `--height` / `--width` to get them:
+- **Defaults** differ. The offline examples auto-detect Edge (from `edge` in the model id /
+  transformer class) and apply the Edge column below, so you can omit `--height` / `--width` /
+  `--guidance-scale` / `--flow-shift` and still get native Edge output:
 
 | | Edge | Nano / Super |
 |---|---|---|
@@ -26,6 +28,10 @@ The small Cosmos3 checkpoint, for constrained deployments. It runs on the shared
 | Video `guidance_scale` | 5.0 | 6.0 |
 | Video `flow_shift` | 3.0 | 10.0 |
 
+> **Important:** these auto-defaults live in the example CLIs
+> (`text_to_video.py` / `image_to_video.py`). Do **not** pass the Nano/Super numbers
+> (720p / `guidance_scale` 6.0 / `flow_shift` 10.0) to Edge — they produce degenerate output.
+
 For the larger checkpoints see [Cosmos3-Nano](./Cosmos3-Nano.md) and [Cosmos3-Super](./Cosmos3-Super.md).
 
 ## Support notes

recipes/cosmos3/Cosmos3-Super.md

@@ -125,3 +125,134 @@ curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
   under the top-level `action` field. Verified on the 64B Super under
   `--cfg-parallel-size 2`: async `policy` returns the predicted action (`[16, 10]`)
   and the rollout video reliably.
+
+## NPU
+
+### 8× Ascend910 (A2, A3)
+
+#### Environment
+
+- OS: Linux
+- Python: 3.10+
+- Driver / runtime: Ascend NPU driver with CANN toolkit
+- Recommended operator library: **mindie-sd** (Ascend high-performance fused
+  operators — enables `adalayernorm` and other fused kernels automatically upon
+  installation)
+- vLLM version: Match the repository requirements for your checkout
+- vLLM-Omni version or commit: Use the commit you are deploying from
+
+A pre-built Docker image is available on
+[Docker Hub](https://hub.docker.com/r/vllm/vllm-omni) and
+[Quay.io](https://quay.io/ascend/vllm-omni). Ensure the image tag matches your
+vLLM-Omni checkout so that NPU-specific code is in sync with the container.
+
+#### Prerequisites
+
+Install the **mindie-sd** operator library to enable Ascend-optimized fused
+operators (`adalayernorm`, etc.):
+
+```bash
+git clone https://gitcode.com/Ascend/MindIE-SD.git && cd MindIE-SD
+
+# Comment out the tik_ops build step (not needed for this use case)
+sed -i 's|^\(\s*\)source ${current_script_dir}/build_tik_ops.sh|\1# source ${current_script_dir}/build_tik_ops.sh|' build/build_ops.sh
+
+python setup.py bdist_wheel
+cd dist
+pip install mindiesd-*.whl
+```
+
+After installation, enable the Laser Attention kernel for significant
+long-sequence speedups:
+
+```bash
+export MINDIE_SD_FA_TYPE=ascend_laser_attention
+```
+
+#### Command
+
+```bash
+export MINDIE_SD_FA_TYPE=ascend_laser_attention
+
+vllm serve nvidia/Cosmos3-Super \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --tensor-parallel-size 8 \
+  --model-class-name Cosmos3OmniDiffusersPipeline \
+  --no-guardrails \
+  --init-timeout 1800
+```
+
+#### Verification
+
+Same requests as the GPU section above — all modes (T2I, T2V, I2V, V2V,
+T2VS, I2VS) work identically on NPU. Quick reference with
+`--no-guardrails`:
+
+```bash
+curl http://localhost:8000/v1/models
+
+# T2I (1024x1024, 50 steps)
+curl -sS -X POST http://localhost:8000/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "nvidia/Cosmos3-Super",
+    "prompt": "A photorealistic red sports car on a city street at golden hour, cinematic lighting.",
+    "size": "1024x1024", "n": 1, "response_format": "b64_json",
+    "num_inference_steps": 50, "guidance_scale": 7.0, "seed": 42
+  }' | python3 -c "import sys,json,base64; open('cosmos3_super_t2i.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
+
+# T2V (1280×720, 189 frames, 35 steps — official params)
+curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Super" -F "prompt=A robot arm is cleaning a plate in the kitchen" \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":false}' \
+  -F "seed=17" -o cosmos3_super_t2v.mp4
+
+# I2V — add an uploaded reference image
+curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Super" -F "prompt=The scene comes to life with smooth, natural motion." \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":false}' \
+  -F "seed=1111" -F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
+  -o cosmos3_super_i2v.mp4
+
+# V2V — add an uploaded reference video. condition_video_keep can be "first" or "last".
+curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Super" -F "prompt=Continue the same scene with smooth natural motion." \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F 'extra_params={"condition_frame_indexes_vision":[0,1],"condition_video_keep":"first","guardrails":false}' \
+  -F "seed=2222" -F "input_reference=@/path/to/reference.mp4;type=video/mp4" \
+  -o cosmos3_super_v2v.mp4
+
+# T2V + sound — add generate_sound/sound_duration (output muxes AAC 48 kHz stereo)
+curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Super" -F "prompt=A robot arm is cleaning a plate in the kitchen" \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F "generate_sound=true" -F "sound_duration=7.875" \
+  -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":false}' \
+  -F "seed=17" -o cosmos3_super_t2vs.mp4
+
+# I2V + sound — reference image with synchronized audio
+curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Super" -F "prompt=The scene comes to life with smooth, natural motion and ambient sound." \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F "generate_sound=true" -F "sound_duration=7.875" \
+  -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":false}' \
+  -F "seed=1111" -F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
+  -o cosmos3_super_i2vs.mp4
+```
+
+#### Notes
+
+- **Parallelism:** `--tensor-parallel-size 8` matches the model's 8 KV heads (`num_key_value_heads: 8`, `num_attention_heads: 64` → GQA). This uses 8 davinci devices (NPU 0–3, both cores each). The remaining NPU 4–7 stay idle.
+- **Model config:** 64 layers × 5120 hidden size × 64 attention heads, ~120 GB transformer on disk (27 shards). Each NPU device loads ~15 GB of model weights.
+- **Peak HBM:** ~22.7 GB per device at startup (bf16 weights). Additional memory is allocated per-generation for KV cache and activations — resolution and frame count drive the peak.
+- **Performance (verified on 8× Ascend910):** T2I 256² / 2 steps / guidance 1.0 → ~1.5 s.
+- **Guardrails** are disabled with `--no-guardrails` (guards are on by default). The gated `nvidia/Cosmos-1.0-Guardrail` model and `cosmos-guardrail` package are not shipped. Add `guardrails: false` in `extra_params` for per-request overrides when the server has guardrails enabled.
+- **Known limitations:** FP8 quantization not yet validated on Ascend NPU. `--enable-layerwise-offload` is available but untested on NPU.

tests/benchmarks/test_daily_omni_pack_modes.py

@@ -0,0 +1,249 @@
+"""Tests for Daily-Omni ``--daily-omni-pack-mode`` packing recipes.
+
+The ``minicpm-interleave`` mode is an accuracy-critical protocol: MiniCPM-o only reaches
+OpenBMB's reported Daily-Omni score when frames and 1s audio segments arrive strictly
+interleaved, with ``max_slice_nums=1`` / ``use_image_id=False`` and an empty MiniCPM
+system message (not the Qwen system prompt). These tests pin that contract so a refactor
+cannot silently regress the score.
+
+vllm stubs are installed by tests/benchmarks/conftest.py before collection.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+# Load the data module directly (bypasses vllm_omni.__init__ heavy imports).
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+_MODULE_PATH = _REPO_ROOT / "vllm_omni" / "benchmarks" / "data_modules" / "daily_omni_dataset.py"
+_MODULE_NAME = "vllm_omni.benchmarks.data_modules.daily_omni_dataset"
+
+if _MODULE_NAME not in sys.modules:
+    _spec = importlib.util.spec_from_file_location(_MODULE_NAME, _MODULE_PATH)
+    _mod = importlib.util.module_from_spec(_spec)
+    sys.modules[_MODULE_NAME] = _mod
+    _spec.loader.exec_module(_mod)
+
+from vllm_omni.benchmarks.data_modules.daily_omni_dataset import (  # noqa: E402
+    DailyOmniDataset,
+    _uniform_sample_indices,
+)
+
+_VIDEO_ID = "vid001"
+_NUM_FRAMES = 5
+
+
+@pytest.fixture
+def video_dir(tmp_path: Path) -> Path:
+    """Official Daily-Omni ``Videos/`` layout with placeholder media files."""
+    per_video = tmp_path / "Videos" / _VIDEO_ID
+    per_video.mkdir(parents=True)
+    (per_video / f"{_VIDEO_ID}_video.mp4").write_bytes(b"fake-mp4")
+    (per_video / f"{_VIDEO_ID}_audio.wav").write_bytes(b"fake-wav")
+    return tmp_path / "Videos"
+
+
+@pytest.fixture
+def qa_json(tmp_path: Path) -> Path:
+    path = tmp_path / "qa.json"
+    path.write_text(
+        json.dumps(
+            [
+                {
+                    "video_id": _VIDEO_ID,
+                    "question": "What happens first?",
+                    "choice": {"A": "a", "B": "b", "C": "c", "D": "d"},
+                    "answer": "A",
+                    "task_type": "AV Event Alignment",
+                    "video_duration": "30s",
+                }
+            ]
+        ),
+        encoding="utf-8",
+    )
+    return path
+
+
+def _make_dataset(qa_json: Path, video_dir: Path, **kwargs) -> DailyOmniDataset:
+    return DailyOmniDataset(
+        qa_json_path=str(qa_json),
+        video_dir=str(video_dir),
+        disable_shuffle=True,
+        **kwargs,
+    )
+
+
+def _stub_extract(monkeypatch, dataset: DailyOmniDataset, counter: list[int]) -> None:
+    """Replace decord/ffmpeg extraction with deterministic in-memory frames + segments."""
+    from PIL import Image
+
+    def fake_extract(video_path, *, audio_path, include_audio, **_kw):
+        counter.append(1)
+        frames = [Image.new("RGB", (16, 16), color=(i, i, i)) for i in range(_NUM_FRAMES)]
+        segments = [np.zeros(1600, dtype=np.float32) for _ in range(_NUM_FRAMES)] if include_audio else []
+        return frames, segments
+
+    monkeypatch.setattr(dataset, "_extract_minicpm_frame_audio_segments", fake_extract)
+
+
+# ---------------------------------------------------------------------------
+# _uniform_sample_indices
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(("n", "k"), [(10, 4), (100, 64), (1000, 64), (7, 7), (3, 5), (1, 1)])
+def test_uniform_sample_indices_matches_openbmb_midpoint(n: int, k: int) -> None:
+    """Pin OpenBMB midpoint-bin sampling (not np.linspace endpoints)."""
+    got = _uniform_sample_indices(n, k)
+    if n <= k:
+        assert got == list(range(n))
+    else:
+        gap = n / k
+        expected = [int(i * gap + gap / 2) for i in range(k)]
+        assert got == expected
+        # Endpoints differ from linspace for typical Daily-Omni downsample sizes.
+        linspace = [int(i) for i in np.linspace(0, n - 1, k, dtype=int)]
+        if n == 1000 and k == 64:
+            assert got[0] == 7 and got[-1] == 992
+            assert linspace[0] == 0 and linspace[-1] == 999
+            assert got != linspace
+
+
+def test_uniform_sample_indices_non_positive_k() -> None:
+    assert _uniform_sample_indices(10, 0) == []
+    assert _uniform_sample_indices(10, -1) == []
+
+
+# ---------------------------------------------------------------------------
+# minicpm-interleave packing
+# ---------------------------------------------------------------------------
+
+
+def test_minicpm_interleave_alternates_image_and_audio(monkeypatch, qa_json, video_dir) -> None:
+    ds = _make_dataset(qa_json, video_dir, pack_mode="minicpm-interleave", input_mode="all")
+    _stub_extract(monkeypatch, ds, [])
+
+    parts, extra, position = ds._compose_daily_omni_multimodal(_VIDEO_ID, None)
+
+    assert position == "first"
+    assert [p["type"] for p in parts] == ["image_url", "audio_url"] * _NUM_FRAMES
+    assert extra["mm_processor_kwargs"] == {
+        "use_audio_in_video": False,
+        "max_slice_nums": 1,
+        "use_image_id": False,
+    }
+
+
+def test_minicpm_interleave_visual_mode_emits_frames_only(monkeypatch, qa_json, video_dir) -> None:
+    ds = _make_dataset(qa_json, video_dir, pack_mode="minicpm-interleave", input_mode="visual")
+    _stub_extract(monkeypatch, ds, [])
+
+    parts, _extra, _position = ds._compose_daily_omni_multimodal(_VIDEO_ID, None)
+
+    assert [p["type"] for p in parts] == ["image_url"] * _NUM_FRAMES
+
+
+def test_minicpm_interleave_writes_file_urls_under_video_dir(monkeypatch, qa_json, video_dir) -> None:
+    ds = _make_dataset(qa_json, video_dir, pack_mode="minicpm-interleave", input_mode="all")
+    _stub_extract(monkeypatch, ds, [])
+
+    parts, _extra, _position = ds._compose_daily_omni_multimodal(_VIDEO_ID, None)
+
+    segment_dir = video_dir / ".minicpm_daily_omni_interleave" / _VIDEO_ID
+    for part in parts:
+        url = part[part["type"]]["url"]
+        assert url.startswith("file://")
+        assert Path(url[len("file://") :]).parent == segment_dir.resolve()
+    assert sorted(p.name for p in segment_dir.glob("frame_*.jpg")) == [f"frame_{i:04d}.jpg" for i in range(_NUM_FRAMES)]
+
+
+def test_minicpm_interleave_inline_mode_emits_data_urls(monkeypatch, qa_json, video_dir) -> None:
+    ds = _make_dataset(
+        qa_json,
+        video_dir,
+        pack_mode="minicpm-interleave",
+        input_mode="all",
+        inline_local_video=True,
+    )
+    _stub_extract(monkeypatch, ds, [])
+
+    parts, _extra, _position = ds._compose_daily_omni_multimodal(_VIDEO_ID, None)
+
+    assert parts[0]["image_url"]["url"].startswith("data:image/jpeg;base64,")
+    assert parts[1]["audio_url"]["url"].startswith("data:audio/wav;base64,")
+    assert not (video_dir / ".minicpm_daily_omni_interleave").exists()
+
+
+def test_minicpm_interleave_caches_file_parts_but_not_inline_payloads(monkeypatch, qa_json, video_dir) -> None:
+    """Inline base64 must not be retained: ~700 Daily-Omni videos would pin GBs of RSS."""
+    file_calls: list[int] = []
+    ds_file = _make_dataset(qa_json, video_dir, pack_mode="minicpm-interleave", input_mode="all")
+    _stub_extract(monkeypatch, ds_file, file_calls)
+    ds_file._compose_daily_omni_multimodal(_VIDEO_ID, None)
+    ds_file._compose_daily_omni_multimodal(_VIDEO_ID, None)
+    assert len(file_calls) == 1
+    assert ds_file._minicpm_interleave_cache
+
+    inline_calls: list[int] = []
+    ds_inline = _make_dataset(
+        qa_json,
+        video_dir,
+        pack_mode="minicpm-interleave",
+        input_mode="all",
+        inline_local_video=True,
+    )
+    _stub_extract(monkeypatch, ds_inline, inline_calls)
+    ds_inline._compose_daily_omni_multimodal(_VIDEO_ID, None)
+    ds_inline._compose_daily_omni_multimodal(_VIDEO_ID, None)
+    assert len(inline_calls) == 2
+    assert not ds_inline._minicpm_interleave_cache
+
+
+def test_minicpm_interleave_missing_local_video_returns_none(qa_json, tmp_path) -> None:
+    empty = tmp_path / "empty_videos"
+    empty.mkdir()
+    ds = _make_dataset(qa_json, empty, pack_mode="minicpm-interleave", input_mode="all")
+
+    parts, extra, _position = ds._compose_daily_omni_multimodal(_VIDEO_ID, None)
+
+    assert parts is None
+    assert extra is None
+
+
+# ---------------------------------------------------------------------------
+# System prompt selection
+# ---------------------------------------------------------------------------
+
+
+def test_minicpm_interleave_keeps_empty_system_message(monkeypatch, qa_json, video_dir) -> None:
+    """Official MiniCPM omni mode still sends role=system with empty content."""
+    ds = _make_dataset(qa_json, video_dir, pack_mode="minicpm-interleave", input_mode="all")
+    _stub_extract(monkeypatch, ds, [])
+    parts, _extra, _position = ds._compose_daily_omni_multimodal(_VIDEO_ID, None)
+
+    messages = ds._build_daily_omni_openai_messages(parts, "q?", {"A": "a"})
+
+    assert [m["role"] for m in messages] == ["system", "user"]
+    assert messages[0]["content"][0]["text"] == ""
+
+
+def test_qwen_pack_mode_keeps_system_message(qa_json, video_dir) -> None:
+    ds = _make_dataset(qa_json, video_dir, pack_mode="qwen", input_mode="all")
+
+    messages = ds._build_daily_omni_openai_messages([], "q?", {"A": "a"})
+
+    assert [m["role"] for m in messages] == ["system", "user"]
+    assert "Qwen" in messages[0]["content"][0]["text"]
+
+
+def test_invalid_pack_mode_rejected(qa_json, video_dir) -> None:
+    with pytest.raises(ValueError, match="pack_mode must be"):
+        _make_dataset(qa_json, video_dir, pack_mode="bogus")

tests/config/test_config_factory.py

@@ -1145,7 +1145,7 @@ class TestResolveScheduler:
 
 class TestDeployConfigLoading:
     def test_load_minicpmo_duplex_deploy_config(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "minicpmo_4_5_duplex.yaml"
+        deploy_path = Path(get_deploy_config_path("minicpmo_4_5_duplex.yaml"))
 
         deploy = load_deploy_config(deploy_path)
         pipeline = resolve_pipeline_config("minicpmo_4_5")
@@ -1160,8 +1160,10 @@ class TestDeployConfigLoading:
         assert [stage.to_omegaconf().session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
         assert [stage.yaml_engine_args["async_scheduling"] for stage in stages] == [False, False, False]
         assert all("Async" not in (stage.scheduler_cls or "") for stage in stages)
+        assert [stage.devices for stage in deploy.stages] == ["0", "0", "0"]
         assert deploy.stages[1].enforce_eager is False
         assert stages[1].yaml_extras["default_sampling_params"]["max_tokens"] == 2048
+        assert stages[1].yaml_extras["default_sampling_params"]["min_tokens"] == 0
         assert stages[1].yaml_extras["default_sampling_params"]["stop_token_ids"] == [1]
 
     @pytest.mark.parametrize(
@@ -1180,7 +1182,7 @@ class TestDeployConfigLoading:
         stage2_devices: str,
         stage1_replicas: int,
     ):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / filename
+        deploy_path = Path(get_deploy_config_path(filename))
 
         deploy = load_deploy_config(deploy_path)
         pipeline = resolve_pipeline_config("minicpmo_4_5")
@@ -1224,15 +1226,54 @@ stages:
         assert {s.yaml_engine_args.get("custom_voice_dir") for s in stages} == {str(custom_voice_dir)}
 
     def test_load_qwen3_omni_moe_deploy_config(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         deploy = load_deploy_config(deploy_path)
         assert len(deploy.stages) == 3
         assert deploy.async_chunk is True
         assert deploy.connectors is not None
         assert deploy.platforms is not None
 
+    def test_qwen3_omni_thinker_only_deploy_config_uses_resolved_pipeline(self):
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe_thinking.yaml"))
+        with patch(
+            "vllm_omni.config.config_factory.get_config",
+            return_value=Q3_OMNI_THINKER_HF_CONFIG,
+        ):
+            pipeline = StageConfigFactory.get_pipeline_config(
+                "Qwen/Qwen3-Omni-30B-A3B-Thinking",
+                trust_remote_code=False,
+                deploy_config_path=str(deploy_path),
+            )
+
+        assert pipeline is not None
+        assert pipeline.model_type == "qwen3_omni_moe_thinker_only"
+
+        deploy = load_deploy_config(deploy_path)
+        stages = merge_pipeline_deploy(pipeline, deploy)
+
+        assert len(stages) == 1
+        thinker = stages[0]
+        assert thinker.model_stage == "thinker"
+        assert thinker.final_output is True
+        assert thinker.final_output_type == "text"
+        assert thinker.worker_type == "ar"
+        assert thinker.scheduler_cls == "vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler"
+        assert thinker.hf_config_name == "thinker_config"
+        assert thinker.is_comprehension is True
+        assert thinker.yaml_runtime["devices"] == "0,1"
+        assert thinker.yaml_engine_args["engine_output_type"] == "text"
+        assert thinker.yaml_engine_args["tensor_parallel_size"] == 2
+        assert thinker.yaml_engine_args["max_num_seqs"] == 1
+        assert thinker.yaml_engine_args["gpu_memory_utilization"] == 0.9
+        assert thinker.yaml_engine_args["enforce_eager"] is True
+        assert thinker.yaml_engine_args["async_scheduling"] is False
+        assert thinker.yaml_engine_args["distributed_executor_backend"] == "mp"
+        assert thinker.yaml_engine_args["enable_prefix_caching"] is True
+        assert thinker.yaml_engine_args["async_chunk"] is False
+        assert thinker.yaml_extras["default_sampling_params"]["detokenize"] is True
+
     def test_load_voxtral_tts_deploy_config_schema_fields(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "voxtral_tts.yaml"
+        deploy_path = Path(get_deploy_config_path("voxtral_tts.yaml"))
         deploy = load_deploy_config(deploy_path)
         assert deploy.stages[0].config_format == "mistral"
         assert deploy.stages[0].load_format == "mistral"
@@ -1242,7 +1283,7 @@ stages:
         )
 
     def test_load_ming_flash_omni_deploy_config_schema_fields(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "ming_flash_omni.yaml"
+        deploy_path = Path(get_deploy_config_path("ming_flash_omni.yaml"))
         deploy = load_deploy_config(deploy_path)
         assert deploy.stages[0].compilation_config == {"pass_config": {"fuse_allreduce_rms": False}}
         assert "compilation_config" not in deploy.stages[0].engine_extras
@@ -1382,7 +1423,7 @@ stages:
         assert not list(stage_config_dir.glob("*.yaml"))
 
     def test_merge_pipeline_deploy(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 
@@ -1409,7 +1450,7 @@ stages:
         )
         assert isinstance(pipeline, PipelineConfig)
 
-        base = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        base = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not base.exists():
             pytest.skip("Deploy config not found")
 
@@ -1829,7 +1870,7 @@ class TestQwen3TTSPipeline:
 
     def test_per_stage_model_arch_flows_through_merge(self, tmp_path):
         """Verify the new ps.model_arch override survives merge_pipeline_deploy."""
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_tts.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_tts.yaml"))
         if not deploy_path.exists():
             pytest.skip("qwen3_tts deploy yaml not found")
 
@@ -1953,7 +1994,7 @@ class TestMingFlashOmniPipeline:
 
     def test_full_yaml_loads_and_merges(self):
         """deploy/ming_flash_omni.yaml parses and merges with the registered pipeline."""
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "ming_flash_omni.yaml"
+        deploy_path = Path(get_deploy_config_path("ming_flash_omni.yaml"))
         if not deploy_path.exists():
             pytest.skip("ming_flash_omni deploy yaml not found")
 
@@ -1973,7 +2014,7 @@ class TestMingFlashOmniPipeline:
 
     def test_tts_yaml_loads_and_merges(self):
         """deploy/ming_flash_omni_tts.yaml parses and routes to the TTS-only pipeline."""
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "ming_flash_omni_tts.yaml"
+        deploy_path = Path(get_deploy_config_path("ming_flash_omni_tts.yaml"))
         if not deploy_path.exists():
             pytest.skip("ming_flash_omni_tts deploy yaml not found")
 
@@ -2011,7 +2052,7 @@ class TestMingFlashOmniPipeline:
 
     def test_thinker_only_yaml_loads_and_merges(self):
         """deploy/ming_flash_omni_thinker_only.yaml parses and routes to the thinker-only pipeline."""
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "ming_flash_omni_thinker_only.yaml"
+        deploy_path = Path(get_deploy_config_path("ming_flash_omni_thinker_only.yaml"))
         if not deploy_path.exists():
             pytest.skip("ming_flash_omni_thinker_only deploy yaml not found")
 
@@ -2071,7 +2112,7 @@ class TestMingFlashOmniPipeline:
 
     def test_image_yaml_loads_and_merges(self):
         """deploy/ming_flash_omni_image.yaml parses and routes to the image pipeline."""
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "ming_flash_omni_image.yaml"
+        deploy_path = Path(get_deploy_config_path("ming_flash_omni_image.yaml"))
         if not deploy_path.exists():
             pytest.skip("ming_flash_omni_image deploy yaml not found")
 
@@ -2133,7 +2174,7 @@ class TestBaseConfigInheritance:
 
     def test_pure_inheritance_overlay(self, tmp_path):
         """An overlay with only ``base_config`` inherits everything."""
-        base = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        base = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not base.exists():
             pytest.skip("Base deploy config not found")
 
@@ -2146,7 +2187,7 @@ class TestBaseConfigInheritance:
 
     def test_single_field_overlay(self, tmp_path):
         """An overlay overriding one stage field merges with the base."""
-        base = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        base = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not base.exists():
             pytest.skip("Base deploy config not found")
 
@@ -2162,8 +2203,45 @@ class TestBaseConfigInheritance:
 class TestPlatformOverrides:
     """Test platform-specific deploy config overrides."""
 
+    def test_qwen3_tts_rocm_disables_code2wav_outer_cudagraph(self):
+        deploy_path = Path(get_deploy_config_path("qwen3_tts.yaml"))
+
+        base = load_deploy_config(deploy_path)
+        assert base.stages[0].enforce_eager is None
+        assert base.stages[1].enforce_eager is False
+
+        rocm = _apply_platform_overrides(base, platform="rocm")
+        assert rocm.stages[0].enforce_eager is None
+        assert rocm.stages[1].enforce_eager is True
+
+    def test_minicpmo_4_5_cuda_caps_talker_kv_cache(self):
+        pipeline = resolve_pipeline_config("minicpmo_4_5")
+        assert isinstance(pipeline, PipelineConfig)
+        deploy_path = Path(get_deploy_config_path("minicpmo_4_5.yaml"))
+
+        cuda = _apply_platform_overrides(load_deploy_config(deploy_path), platform="cuda")
+        cuda_stages = merge_pipeline_deploy(pipeline, cuda)
+        assert cuda_stages[1].yaml_engine_args["kv_cache_memory_bytes"] == 2 * 1024**3
+
+        # The CUDA memory budget must not leak into the existing NPU profile.
+        npu = _apply_platform_overrides(load_deploy_config(deploy_path), platform="npu")
+        npu_stages = merge_pipeline_deploy(pipeline, npu)
+        assert "kv_cache_memory_bytes" not in npu_stages[1].yaml_engine_args
+
+        # Multi-GPU replica profiles place Talker on dedicated devices and
+        # preserve their previous automatic KV-cache sizing.
+        for filename in (
+            "minicpmo_4_5_3gpu_stage1_replicas.yaml",
+            "minicpmo_4_5_4gpu_stage1_replicas.yaml",
+        ):
+            replica = _apply_platform_overrides(
+                load_deploy_config(Path(get_deploy_config_path(filename))), platform="cuda"
+            )
+            replica_stages = merge_pipeline_deploy(pipeline, replica)
+            assert replica_stages[1].yaml_engine_args["kv_cache_memory_bytes"] is None
+
     def test_npu_overrides(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 
@@ -2177,7 +2255,7 @@ class TestPlatformOverrides:
         assert deploy.stages[2].enforce_eager is False
 
     def test_xpu_overrides(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 
@@ -2189,7 +2267,7 @@ class TestPlatformOverrides:
         assert deploy.stages[0].engine_extras.get("max_cudagraph_capture_size") == 0
 
     def test_unknown_platform_noop(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 
@@ -2234,7 +2312,7 @@ class TestCLIOverrideFlow:
     """Test --stage-overrides JSON merge into StageConfig."""
 
     def test_stage_overrides_merge(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 
@@ -2252,7 +2330,7 @@ class TestCLIOverrideFlow:
         assert stages[0].runtime_overrides["gpu_memory_utilization"] == 0.5
 
     def test_global_override_applies_to_all(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 
@@ -2272,7 +2350,7 @@ class TestCLIOverrideFlow:
 
 class TestAuraOmniDeploy:
     def test_aura_omni_deploy_forces_pipeline_override(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "aura_omni.yaml"
+        deploy_path = Path(get_deploy_config_path("aura_omni.yaml"))
         deploy = load_deploy_config(deploy_path)
 
         assert deploy.pipeline == "aura_omni"
@@ -2283,7 +2361,7 @@ class TestAuraOmniDeploy:
         stages, _ = StageConfigFactory._create_legacy_from_registry(
             pipeline_cfg,
             cli_overrides={},
-            deploy_config_path=str(Path(__file__).parent.parent / "vllm_omni" / "deploy" / "aura_omni.yaml"),
+            deploy_config_path=get_deploy_config_path("aura_omni.yaml"),
         )
 
         assert [stage.model_stage for stage in stages] == [
@@ -2597,7 +2675,12 @@ class TestSentinelDefaultPrecedence:
         from vllm_omni.config.stage_config import DeployConfig, merge_pipeline_deploy
 
         pipeline_path = (
-            Path(__file__).parent.parent / "vllm_omni" / "model_executor" / "models" / "qwen3_omni" / "pipeline.py"
+            Path(__file__).parent.parent.parent
+            / "vllm_omni"
+            / "model_executor"
+            / "models"
+            / "qwen3_omni"
+            / "pipeline.py"
         )
         pipeline = runpy.run_path(str(pipeline_path))["QWEN3_OMNI_PIPELINE"]
 
@@ -2667,7 +2750,7 @@ class TestSamplingConstraintsPrecedence:
     """Test that pipeline sampling_constraints override deploy defaults."""
 
     def test_constraints_win(self):
-        deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
+        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
         if not deploy_path.exists():
             pytest.skip("Deploy config not found")
 

tests/config/test_omni_config.py

@@ -440,6 +440,7 @@ def test_sub_config_fields_match_rfc_scopes():
         "ulysses_mode",
         "cfg_parallel_size",
         "vae_patch_parallel_size",
+        "text_encoder_tp_size",
         "vae_parallel_mode",
         "use_hsdp",
         "mask_sp_padding",
@@ -577,9 +578,11 @@ def test_from_pipeline_config_derives_sequence_parallel_size_from_allgather_degr
     assert stage.parallel_config.world_size == 2
 
 
-def test_diffusion_parallel_config_rejects_cfg_parallel_size_outside_current_bound():
-    with pytest.raises(ValidationError):
-        OmniStageDiffusionParallelConfig(cfg_parallel_size=4)
+def test_diffusion_parallel_config_accepts_four_way_guidance_parallelism():
+    cfg = OmniStageDiffusionParallelConfig(cfg_parallel_size=4)
+
+    assert cfg.cfg_parallel_size == 4
+    assert cfg.world_size == 4
 
 
 def test_diffusion_parallel_config_rejects_allgather_with_ulysses_or_ring():

tests/conftest.py

@@ -3,6 +3,7 @@ Root pytest entrypoint for the vLLM-Omni test suite.
 
 - `tests/conftest.py` stays thin: plugin registration + compatibility re-exports.
 - Importable utilities live under `tests/helpers/`.
+- Helper unit tests live under `tests/helpers/tests/`.
 - Fixtures live under `tests/helpers/fixtures/` and are loaded via `pytest_plugins`.
 """
 

tests/core/sched/test_omni_ar_scheduler_free_request_cleanup.py

@@ -0,0 +1,74 @@
+"""Regression tests for vllm-project/vllm-omni#5349 (P1): normal completion
+goes through _free_request(), not finish_requests() (the external
+abort/cancel entry point). Without a cleanup_receiver() call there,
+_active_streams never releases on ordinary completion, so the bounded-K
+window fills with stale entries after K completions.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+# Imports must run in this order: vllm_omni applies patches to vllm.v1.request before
+# Request / RequestStatus are bound in this module. Ruff isort would reorder them.
+# isort: off
+import vllm_omni  # noqa: F401 - import for side effects (patch vLLM)
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+
+# isort: on
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+def _make_scheduler(*, chunk_transfer_adapter=None) -> OmniARScheduler:
+    """Minimal OmniARScheduler exercising _free_request()'s no-KV-transfer
+    happy path."""
+    sched = OmniARScheduler.__new__(OmniARScheduler)
+    sched._omits_kv_transfer_cache = {}
+    sched._connector_finished = lambda request: (False, None)
+    sched.encoder_cache_manager = MagicMock()
+    sched.finished_req_ids = set()
+    sched._new_prompt_len_snapshot = {}
+    sched.finished_req_ids_dict = None
+    sched._should_transfer_kv_for_request = lambda req_id: False
+    sched._free_blocks = MagicMock()
+    sched._free_input_coordinator_request = MagicMock()
+    sched.chunk_transfer_adapter = chunk_transfer_adapter
+    return sched
+
+
+class _FakeFinishedRequest:
+    """Not a SimpleNamespace: SimpleNamespace defines __eq__, which makes it
+    unhashable, and _free_request() puts the request in a set."""
+
+    def __init__(self, request_id: str) -> None:
+        self.request_id = request_id
+
+    def is_finished(self) -> bool:
+        return True
+
+
+def _make_finished_request(request_id: str = "req-1"):
+    return _FakeFinishedRequest(request_id)
+
+
+def test_free_request_releases_chunk_transfer_adapter_receiver_state():
+    adapter = MagicMock()
+    sched = _make_scheduler(chunk_transfer_adapter=adapter)
+    request = _make_finished_request("req-1")
+
+    sched._free_request(request)
+
+    adapter.cleanup_receiver.assert_called_once_with("req-1")
+
+
+def test_free_request_is_safe_without_a_chunk_transfer_adapter():
+    """Most stages have no chunk transfer adapter configured."""
+    sched = _make_scheduler(chunk_transfer_adapter=None)
+    request = _make_finished_request("req-1")
+
+    sched._free_request(request)  # must not raise
+
+    sched._free_input_coordinator_request.assert_called_once_with("req-1")

tests/core/sched/test_omni_ar_scheduler_kv_transfer.py

@@ -0,0 +1,57 @@
+from types import SimpleNamespace
+
+import pytest
+from vllm import SamplingParams
+from vllm.v1.engine import EngineCoreRequest
+
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+from vllm_omni.engine.async_engine_utils import apply_omni_final_stage_metadata
+from vllm_omni.engine.serialization import deserialize_additional_information
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+def _make_engine_request() -> EngineCoreRequest:
+    return EngineCoreRequest(
+        request_id="req",
+        prompt_token_ids=[1],
+        mm_features=None,
+        sampling_params=SamplingParams(max_tokens=1),
+        pooling_params=None,
+        arrival_time=0.0,
+        lora_request=None,
+        cache_salt=None,
+        data_parallel_rank=None,
+    )
+
+
+def _request_omits_kv_transfer(*, force_kv_transfer: bool) -> tuple[bool, dict]:
+    tagged = apply_omni_final_stage_metadata(
+        _make_engine_request(),
+        final_stage_id=0,
+        force_kv_transfer=force_kv_transfer,
+    )
+    scheduler = OmniARScheduler.__new__(OmniARScheduler)
+    scheduler._omits_kv_transfer_cache = {}
+    request = SimpleNamespace(
+        request_id="req",
+        additional_information=tagged.additional_information,
+    )
+    result = scheduler._request_omits_kv_transfer_to_next_stage(request)
+    metadata = deserialize_additional_information(tagged.additional_information)
+    return result, metadata
+
+
+def test_stage_zero_request_omits_kv_transfer():
+    omits_transfer, metadata = _request_omits_kv_transfer(force_kv_transfer=False)
+
+    assert omits_transfer
+    assert "omni_force_kv_transfer" not in metadata
+
+
+def test_cfg_companion_forces_kv_transfer_without_downstream_payload():
+    omits_transfer, metadata = _request_omits_kv_transfer(force_kv_transfer=True)
+
+    assert not omits_transfer
+    assert metadata["omni_final_stage_id"] == 0
+    assert metadata["omni_force_kv_transfer"] is True

tests/core/sched/test_omni_ar_scheduler_logprobs.py

@@ -3,7 +3,8 @@
 from __future__ import annotations
 
 from collections import defaultdict
-from types import SimpleNamespace
+from collections.abc import Callable
+from types import MethodType, SimpleNamespace
 
 import numpy as np
 import pytest
@@ -12,9 +13,22 @@ from vllm.v1.outputs import LogprobsLists
 from vllm.v1.request import RequestStatus
 
 from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler, _slice_sampled_logprobs
+from vllm_omni.core.sched.omni_scheduler_mixin import OmniSchedulerMixin
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
+# Shared update_from_output helpers moved onto OmniSchedulerMixin; bind them on
+# the lightweight stub so direct method calls still exercise the real paths.
+_MIXIN_UPDATE_HELPERS = (
+    "_remove_stopped_requests_from_queues",
+    "_handle_failed_kv_load_outputs",
+    "_aggregate_kv_connector_stats",
+    "_publish_kv_cache_events",
+    "_attach_finished_request_sets",
+    "_attach_scheduler_stats",
+    "_capture_omni_connector_output",
+)
+
 
 def _logprobs(
     token_ids: list[list[int]],
@@ -76,6 +90,7 @@ class _Request:
         self.status = RequestStatus.RUNNING
         self.sampling_params = SimpleNamespace(num_logprobs=1)
         self.num_computed_tokens = 0
+        self.num_in_flight_tokens = 0
         self.num_output_placeholders = 0
         self.has_encoder_inputs = False
         self.pooling_params = None
@@ -104,17 +119,14 @@ class _RequestQueue(list):
                 self.remove(request)
 
 
-@pytest.mark.skip(reason="#issue 5484")
-def test_invalid_logprobs_finish_only_the_affected_scheduler_request() -> None:
-    """A bad runner response must not bring down unrelated batch requests."""
-    bad = _Request("bad")
-    good = _Request("good")
+def _make_scheduler_stub(requests: list[_Request]) -> SimpleNamespace:
+    """Build a minimal stub for OmniARScheduler.update_from_output()."""
     scheduler = SimpleNamespace(
         perf_metrics=None,
         connector=None,
         chunk_transfer_adapter=None,
-        requests={bad.request_id: bad, good.request_id: good},
-        running=[bad, good],
+        requests={request.request_id: request for request in requests},
+        running=list(requests),
         waiting=_RequestQueue(),
         skipped_waiting=_RequestQueue(),
         structured_output_manager=SimpleNamespace(should_advance=lambda _request: False),
@@ -127,26 +139,98 @@ def test_invalid_logprobs_finish_only_the_affected_scheduler_request() -> None:
         finished_req_ids_dict=defaultdict(set),
         kv_cache_manager=SimpleNamespace(take_events=lambda: None),
         kv_event_publisher=SimpleNamespace(publish=lambda _events: None),
+        recompute_kv_load_failures=False,
     )
+    for name in _MIXIN_UPDATE_HELPERS:
+        setattr(scheduler, name, MethodType(getattr(OmniSchedulerMixin, name), scheduler))
+    scheduler._cleanup_kv_tracking = MethodType(OmniARScheduler._cleanup_kv_tracking, scheduler)
+    scheduler.make_spec_decoding_stats = lambda *args, **kwargs: None
+    scheduler.make_stats = lambda *args, **kwargs: None
+    return scheduler
+
+
+def _bind_request_lifecycle(
+    scheduler: SimpleNamespace,
+    *,
+    update_request: Callable,
+    handle_stopped: Callable | None = None,
+) -> None:
+    def free_request(request):
+        scheduler.requests.pop(request.request_id, None)
+        scheduler.finished_req_ids.add(request.request_id)
+        scheduler.finished_req_ids_dict[request.client_index].add(request.request_id)
+        # vLLM 0.26 contract: (kv_xfer_params, ec_xfer_params)
+        return None, None
+
+    scheduler._update_request_with_output = update_request
+    scheduler._process_kv_transfer_trigger = lambda _request, _tokens: False
+    scheduler._handle_stopped_request = handle_stopped or (lambda _request: True)
+    scheduler._free_request = free_request
+
+
+def test_mid_step_stop_trims_logprob_rows_with_token_ids() -> None:
+    """Spec decode can sample tokens past a stop; the trailing tokens are
+    trimmed by _update_request_with_output, and the emitted logprob rows must
+    be trimmed with them (upstream slices logprobs after the trim)."""
+    request = _Request("req")
+    request.num_computed_tokens = 8
+    scheduler = _make_scheduler_stub([request])
+
+    def update_request_trimming(req, token_ids):
+        # Mirror vllm Scheduler._update_request_with_output: the first sampled
+        # token hits EOS, so the spec-accepted tokens behind it are trimmed
+        # in place and the request stops.
+        del token_ids[1:]
+        req.status = RequestStatus.FINISHED_STOPPED
+        return token_ids, True
+
+    _bind_request_lifecycle(scheduler, update_request=update_request_trimming)
+
+    scheduler_output = SimpleNamespace(
+        num_scheduled_tokens={"req": 3},
+        scheduled_spec_decode_tokens={"req": [8, 9]},
+        num_invalid_spec_tokens=0,
+    )
+    model_runner_output = SimpleNamespace(
+        sampled_token_ids=[[7, 8, 9]],
+        logprobs=_logprobs(
+            [[7, 1], [8, 2], [9, 3]],
+            [[-0.7, -1.0], [-0.8, -1.1], [-0.9, -1.2]],
+            cumulative_rows=[0],
+        ),
+        prompt_logprobs_dict={},
+        pooler_output=None,
+        num_nans_in_logits=None,
+        kv_connector_output=None,
+        cudagraph_stats=None,
+        req_id_to_index={"req": 0},
+        routed_experts=None,
+    )
+
+    outputs = OmniARScheduler.update_from_output(scheduler, scheduler_output, model_runner_output)
+    (output,) = outputs[0].outputs
+
+    assert output.new_token_ids == [7]
+    assert output.finish_reason is FinishReason.STOP
+    token_rows = np.asarray(output.new_logprobs.logprob_token_ids)
+    value_rows = np.asarray(output.new_logprobs.logprobs)
+    assert token_rows.shape[0] == len(output.new_token_ids)
+    assert value_rows.shape[0] == len(output.new_token_ids)
+    np.testing.assert_array_equal(token_rows[:, 0], [7])
+
+
+def test_invalid_logprobs_finish_only_the_affected_scheduler_request() -> None:
+    """A bad runner response must not bring down unrelated batch requests."""
+    bad = _Request("bad")
+    good = _Request("good")
+    scheduler = _make_scheduler_stub([bad, good])
     update_calls: list[str] = []
 
     def update_request(request, token_ids):
         update_calls.append(request.request_id)
         return token_ids, False
 
-    def free_request(request):
-        scheduler.requests.pop(request.request_id, None)
-        scheduler.finished_req_ids.add(request.request_id)
-        scheduler.finished_req_ids_dict[request.client_index].add(request.request_id)
-        return None
-
-    scheduler._update_request_with_output = update_request
-    scheduler._process_kv_transfer_trigger = lambda _request, _tokens: False
-    scheduler._handle_stopped_request = lambda _request: True
-    scheduler._free_request = free_request
-    scheduler.make_spec_decoding_stats = lambda *args, **kwargs: None
-    scheduler.make_stats = lambda *args, **kwargs: None
-    scheduler._capture_omni_connector_output = lambda _output: None
+    _bind_request_lifecycle(scheduler, update_request=update_request)
 
     scheduler_output = SimpleNamespace(
         num_scheduled_tokens={"bad": 1, "good": 1},

tests/core/sched/test_omni_ar_scheduler_streaming.py

@@ -262,9 +262,52 @@ def test_explicit_streaming_payload_replaces_placeholder_prompt() -> None:
         "tts_token_ids": [10, 20],
         "meta": {"replace_streaming_prompt": True},
     }
+    update.model_intermediate_buffer = {
+        "ids": {"tts": [41, 42, 99]},
+        "meta": {"turn_eos_token_id": 99},
+    }
 
     sched._update_request_as_session(session, update)
 
     assert session.prompt_token_ids == [10, 20]
     assert session.additional_information == update.additional_information
+    assert session.model_intermediate_buffer == {
+        "ids": {"tts": [41, 42, 99]},
+        "meta": {"turn_eos_token_id": 99},
+    }
+    assert session.status == RequestStatus.WAITING
+
+
+def test_model_intermediate_streaming_payload_replaces_computed_prompt() -> None:
+    sched = _make_scheduler(stage_id=1)
+    sched.chunk_transfer_adapter = SimpleNamespace(
+        receives_chunks=False,
+        segment_finished_requests=set(),
+    )
+    session = _make_request()
+    session.status = RequestStatus.WAITING_FOR_STREAMING_REQ
+    session.prompt_token_ids = [0] * 59
+    session._all_token_ids.clear()
+    session._all_token_ids.extend(session.prompt_token_ids)
+    session.num_prompt_tokens = 59
+    session.num_computed_tokens = 59
+    update = _make_update([0] * 10)
+    update.additional_information = None
+    update.model_intermediate_buffer = {
+        "ids": {"tts": list(range(8))},
+        "hidden_states": {"tts": [[0.0]] * 8},
+        "meta": {
+            "replace_streaming_prompt": True,
+            "next_stage_prompt_len": 10,
+        },
+    }
+
+    sched._update_request_as_session(session, update)
+
+    assert session.prompt_token_ids == [0] * 10
+    assert list(session._all_token_ids) == [0] * 10
+    assert session.num_prompt_tokens == 10
+    assert session.num_computed_tokens == 0
+    assert session.additional_information is None
+    assert session.model_intermediate_buffer == update.model_intermediate_buffer
     assert session.status == RequestStatus.WAITING

tests/core/sched/test_omni_generation_scheduler_update_session.py

@@ -21,6 +21,7 @@ import vllm_omni  # noqa: F401 - import for side effects (patch vLLM)
 from vllm.sampling_params import SamplingParams
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.engine import EngineCoreEventType
+from vllm.v1.metrics.stats import PrefillStats
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.v1.request import Request, RequestStatus, StreamingUpdate
 from vllm_omni.core.sched.omni_generation_scheduler import OmniGenerationScheduler
@@ -54,6 +55,18 @@ class _SchedulerStub(OmniGenerationScheduler):
         raise AssertionError("unexpected enqueue for skipped_waiting miss")
 
 
+class _AsyncChunkStopSchedulerStub(OmniGenerationScheduler):
+    def __init__(self) -> None:
+        self.num_waiting_for_streaming_input = 0
+        self.chunk_transfer_adapter = SimpleNamespace(receives_chunks=True)
+        self.enqueued: list[Request] = []
+        self.enqueued_statuses: list[RequestStatus] = []
+
+    def _enqueue_waiting_request(self, session: Request) -> None:
+        self.enqueued.append(session)
+        self.enqueued_statuses.append(session.status)
+
+
 def _make_request(**kwargs) -> Request:
     sp = SamplingParams(max_tokens=8)
     defaults = dict(
@@ -81,6 +94,17 @@ def _make_update(**kwargs) -> StreamingUpdate:
     return StreamingUpdate(**defaults)
 
 
+def test_generation_scheduler_records_prefill_stats_for_metrics() -> None:
+    request = _make_request(prompt_token_ids=[1, 2, 3, 4])
+    request.prefill_stats = PrefillStats()
+
+    OmniGenerationScheduler._record_prefill_stats(request)
+
+    assert request.prefill_stats.num_prompt_tokens == 4
+    assert request.prefill_stats.num_computed_tokens == 4
+    assert request.prefill_stats.num_cached_tokens == 0
+
+
 def test_resumable_generation_stop_marks_segment_boundary() -> None:
     session = _make_request(request_id="req-generation-segment")
     session.status = RequestStatus.RUNNING
@@ -133,6 +157,21 @@ def test_resumable_generation_stop_marks_segment_boundary() -> None:
     assert output.is_segment_finished is True
 
 
+def test_async_chunk_resumable_stop_rearms_connector_polling() -> None:
+    sched = _AsyncChunkStopSchedulerStub()
+    session = _make_request(request_id="req-async-chunk-next-segment")
+    session.status = RequestStatus.FINISHED_STOPPED
+    session.resumable = True
+
+    finished = sched._handle_stopped_request(session)
+
+    assert finished is False
+    assert sched.enqueued == [session]
+    assert sched.enqueued_statuses == [RequestStatus.WAITING]
+    assert session.status == RequestStatus.WAITING
+    assert sched.num_waiting_for_streaming_input == 0
+
+
 class TestReplaceSessionWithStreamingUpdate:
     def test_resets_tokens_and_prompt_from_update(self) -> None:
         sched = _SchedulerStub()