[CI][MiniCPM-o] Add NPU Video-MME accuracy nightly as a standalone job (#6596)

Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>

.buildkite/amd/scripts/bootstrap-amd-omni.sh

@@ -6,6 +6,10 @@
 
 set -euo pipefail
 
+# minicpm-challenge branch: skip entire AMD CI (no pipeline upload => no image build / tests).
+echo "--- Skipping AMD CI on this branch (including image build)"
+exit 0
+
 source .buildkite/common/scripts/resolve_skip_ci.sh
 
 if [[ -z "${RUN_ALL:-}" ]]; then

.buildkite/common/scripts/skip_ci.py

@@ -282,7 +282,7 @@ def _git(*args: str) -> subprocess.CompletedProcess[str]:
 
 
 def _resolve_diff_range() -> str | None:
-    """Return a git diff range for PR or main builds, or None when unavailable."""
+    """Return a git diff range for PR or push-branch builds, or None when unavailable."""
     is_pr = os.environ.get("BUILDKITE_PULL_REQUEST", "false") != "false" and os.environ.get(
         "BUILDKITE_PULL_REQUEST", ""
     )
@@ -302,13 +302,15 @@ def _resolve_diff_range() -> str | None:
                 return None
         return f"{base_ref}...{commit}"
 
-    if os.environ.get("BUILDKITE_BRANCH", "") == "main":
+    # main: ready/nightly/weekly push CI; minicpm-challenge: merge (L3) push CI
+    branch = os.environ.get("BUILDKITE_BRANCH", "")
+    if branch in ("main", "minicpm-challenge"):
         if _git("rev-parse", "--verify", f"{commit}^").returncode != 0:
-            _log("main commit has no parent; using safe defaults")
+            _log(f"{branch} commit has no parent; using safe defaults")
             return None
         return f"{commit}^..{commit}"
 
-    _log("not PR/main build; using safe defaults")
+    _log("not PR/main/minicpm-challenge build; using safe defaults")
     return None
 
 

.buildkite/common/scripts/upload_pipeline.py

@@ -59,11 +59,12 @@ E2E_GROUP_MARKER = "E2E Test"
 CI_MIRROR_HARDWARES_PATH = ROOT / ".buildkite/common/ci_mirror_hardwares.yml"
 
 CUDA_NIGHTLY_ONLY = (
-    '(build.pull_request.labels includes "nightly-test") || (build.branch == "main" && build.env("NIGHTLY") == "1")'
+    '(build.pull_request.labels includes "nightly-test") || '
+    '(build.branch == "minicpm-challenge" && build.env("NIGHTLY") == "1")'
 )
 NPU_NIGHTLY_ONLY = (
-    '(build.branch == "main" && build.env("NIGHTLY") == "1") || '
-    '(build.branch != "main" && ('
+    '(build.branch == "minicpm-challenge" && build.env("NIGHTLY") == "1") || '
+    '(build.branch != "minicpm-challenge" && ('
     'build.pull_request.labels includes "nightly-test" || '
     'build.pull_request.labels includes "omni-test" || '
     'build.pull_request.labels includes "tts-test" || '
@@ -103,7 +104,8 @@ def _format_bootstrap_if(expr: str) -> str:
 
 def _compute_bootstrap_if_exprs(*, decision, platform: str) -> dict[str, str]:
     disabled = "false"
-    nightly_main = 'build.branch == "main" && build.env("NIGHTLY") == "1"'
+    # L4 nightly scheduled push CI runs on minicpm-challenge (not main).
+    nightly_main = 'build.branch == "minicpm-challenge" && build.env("NIGHTLY") == "1"'
     nightly_only = NPU_NIGHTLY_ONLY if platform == "npu" else CUDA_NIGHTLY_ONLY
 
     if platform == "npu":
@@ -118,11 +120,12 @@ def _compute_bootstrap_if_exprs(*, decision, platform: str) -> dict[str, str]:
         merge_base = disabled
     else:
         ready_pr = 'build.branch != "main" && build.pull_request.labels includes "ready"'
-        merge_main = 'build.branch == "main" && build.env("NIGHTLY") != "1" && build.env("WEEKLY") != "1"'
-        merge_pr = 'build.branch != "main" && build.pull_request.labels includes "merge-test"'
+        # L3 merge push CI runs on minicpm-challenge (not main).
+        merge_main = 'build.branch == "minicpm-challenge" && build.env("NIGHTLY") != "1" && build.env("WEEKLY") != "1"'
+        merge_pr = 'build.branch != "minicpm-challenge" && build.pull_request.labels includes "merge-test"'
         nightly_label_if = (
-            '(build.branch == "main" && build.env("NIGHTLY") == "1") || '
-            '(build.branch != "main" && ('
+            '(build.branch == "minicpm-challenge" && build.env("NIGHTLY") == "1") || '
+            '(build.branch != "minicpm-challenge" && ('
             'build.pull_request.labels includes "nightly-test" || '
             'build.pull_request.labels includes "omni-test" || '
             'build.pull_request.labels includes "tts-test" || '
@@ -140,8 +143,8 @@ def _compute_bootstrap_if_exprs(*, decision, platform: str) -> dict[str, str]:
     weekly_main = 'build.branch == "main" && build.env("WEEKLY") == "1"'
 
     if decision.skip_all:
-        # Docs / skip-mark only: no PR-label escape hatch. Main scheduled
-        # NIGHTLY=1 still runs L4 + L2/L3 (--e2e); WEEKLY=1 still runs L5.
+        # Docs / skip-mark only: no PR-label escape hatch. Scheduled
+        # NIGHTLY=1 (minicpm-challenge) still runs L4 + L2/L3 (--e2e); WEEKLY=1 (main) still runs L5.
         image_expr = f"({nightly_main}) || ({weekly_main})" if platform == "cuda" else nightly_main
         ready_expr = nightly_main
         merge_expr = nightly_main if platform == "cuda" else disabled

.buildkite/cuda/bootstrap-upload-steps.yml

@@ -8,35 +8,35 @@ steps:
     agents:
       queue: "cpu_queue_premerge"
 
-  # L2 Test — PR ready label (diff-aware) or main+NIGHTLY=1 (--e2e, full E2E group)
+  # L2 Test — PR ready label (diff-aware) or minicpm-challenge+NIGHTLY=1 (--e2e, full E2E group)
   - label: "Upload Ready Pipeline"
     depends_on: image-build
     key: upload-ready-pipeline
     commands:
       - |
         EXTRA=""
-        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "minicpm-challenge" ]; then
           EXTRA="--e2e"
         fi
         python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-ready.yml
     agents:
       queue: "cpu_queue_premerge"
 
-  # L3 Test — PR merge-test label (diff-aware) or main+NIGHTLY=1 (--e2e, full E2E group)
+  # L3 Test — PR merge-test label (diff-aware) or minicpm-challenge+NIGHTLY=1 (--e2e, full E2E group)
   - label: "Upload Merge Pipeline"
     depends_on: image-build
     key: upload-merge-pipeline
     commands:
       - |
         EXTRA=""
-        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "minicpm-challenge" ]; then
           EXTRA="--e2e"
         fi
         python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-merge.yml
     agents:
       queue: "cpu_queue_premerge"
 
-  # L4 Test — main+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
+  # L4 Test — minicpm-challenge+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
   - label: "Upload Nightly Pipeline"
     depends_on: image-build
     key: upload-nightly-pipeline

.buildkite/cuda/test-merge.yml

@@ -18,257 +18,227 @@ steps:
           - "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"
         mirror_hardwares: l4_1
 
-  - group: ":card_index_dividers: Diffusion Test"
-    depends_on: upload-merge-pipeline
-    steps:
-      - label: "Diffusion · LoRA Test"
-        timeout_in_minutes: 30
-        commands:
-          - pytest -sv tests/diffusion/lora/ -m 'advanced_model and cuda' --run-level "advanced_model"
-        mirror_hardwares: l4_1
+  # - group: ":card_index_dividers: Diffusion Test"
+  #   depends_on: upload-merge-pipeline
+  #   steps:
+  #     - label: "Diffusion · LoRA Test"
+  #       timeout_in_minutes: 30
+  #       commands:
+  #         - 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: "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
-    commands:
-      - |
-        timeout 55m bash -c "
-          set -e
-          export VLLM_IMAGE_FETCH_TIMEOUT=60
-          pytest -s -v tests/distributed -m 'advanced_model and cuda and H100' --run-level 'advanced_model'
-        "
-    mirror_hardwares: h100_1
+  # - label: "Distributed Test"
+  #   depends_on: upload-merge-pipeline
+  #   timeout_in_minutes: 60
+  #   commands:
+  #     - |
+  #       timeout 55m bash -c "
+  #         set -e
+  #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+  #         pytest -s -v tests/distributed -m 'advanced_model and cuda and H100' --run-level 'advanced_model'
+  #       "
+  #   mirror_hardwares: h100_1
 
-  - group: ":card_index_dividers: Entrypoints Test"
-    depends_on: upload-merge-pipeline
-    steps:
-      - label: "Entrypoint Test with H100"
-        timeout_in_minutes: 60
-        commands:
-          - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and H100" --run-level "advanced_model"
-        mirror_hardwares: h100_2
+  # - group: ":card_index_dividers: Entrypoints Test"
+  #   depends_on: upload-merge-pipeline
+  #   steps:
+  #     - label: "Entrypoint Test with H100"
+  #       timeout_in_minutes: 60
+  #       commands:
+  #         - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and H100" --run-level "advanced_model"
+  #       mirror_hardwares: h100_2
 
-      - label: "Entrypoint Test with L4"
-        timeout_in_minutes: 30
-        commands:
-          - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and L4" --run-level "advanced_model"
-        mirror_hardwares: l4_1
+  #     - label: "Entrypoint Test with L4"
+  #       timeout_in_minutes: 30
+  #       commands:
+  #         - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and L4" --run-level "advanced_model"
+  #       mirror_hardwares: l4_1
 
-  - group: ":card_index_dividers: E2E Test"
-    depends_on: upload-merge-pipeline
-    steps:
-      - label: "TTS · Qwen3-TTS CustomVoice Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
-          - tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
+  # - group: ":card_index_dividers: E2E Test"
+  #   depends_on: upload-merge-pipeline
+  #   steps:
+  #     - label: "TTS · Qwen3-TTS CustomVoice Test"
+  #       source_file_dependencies:
+  #         - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
+  #         - tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
+  #         - vllm_omni/model_executor/models/qwen3_tts/
+  #         - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+  #         - vllm_omni/deploy/qwen3_tts.yaml
+  #       commands:
+  #         - |
+  #           timeout 40m bash -c "
+  #             export VLLM_LOGGING_LEVEL=DEBUG
+  #             export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+  #             pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+  #           "
+  #       mirror_hardwares: l4_1
 
-      - label: "TTS · Qwen3-TTS Base Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_base.py
-          - tests/e2e/offline_inference/test_qwen3_tts_base.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
+  #     - label: "TTS · Qwen3-TTS Base Test"
+  #       source_file_dependencies:
+  #         - tests/e2e/online_serving/test_qwen3_tts_base.py
+  #         - tests/e2e/offline_inference/test_qwen3_tts_base.py
+  #         - vllm_omni/model_executor/models/qwen3_tts/
+  #         - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+  #         - vllm_omni/deploy/qwen3_tts.yaml
+  #       commands:
+  #         - |
+  #           timeout 40m bash -c "
+  #             export VLLM_LOGGING_LEVEL=DEBUG
+  #             export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+  #             pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+  #           "
+  #       mirror_hardwares: l4_1
 
-      - label: "TTS · Higgs-Audio-V3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v3.py
-          - vllm_omni/model_executor/models/higgs_audio_v3/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
-          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'advanced_model' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
+  #     - label: "TTS · Higgs-Audio-V3 Test"
+  #       source_file_dependencies:
+  #         - tests/e2e/online_serving/test_higgs_audio_v3.py
+  #         - vllm_omni/model_executor/models/higgs_audio_v3/
+  #         - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
+  #         - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
+  #       commands:
+  #         - |
+  #           timeout 20m bash -c "
+  #             export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  #             export VLLM_USE_DEEP_GEMM=0
+  #             export VLLM_MOE_USE_DEEP_GEMM=0
+  #             pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'advanced_model' --run-level 'advanced_model'
+  #           "
+  #       mirror_hardwares: l4_1
 
-      - label: "Omni · Qwen3-Omni Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_qwen3_omni.py
-          - tests/e2e/online_serving/test_qwen3_omni.py
-          - vllm_omni/model_executor/models/qwen3_omni/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
-          - vllm_omni/deploy/qwen3_omni_moe.yaml
-        timeout_in_minutes: 50
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        mirror_hardwares: h100_2
+  #     - label: "Omni · Qwen3-Omni Test"
+  #       source_file_dependencies:
+  #         - tests/e2e/offline_inference/test_qwen3_omni.py
+  #         - tests/e2e/online_serving/test_qwen3_omni.py
+  #         - vllm_omni/model_executor/models/qwen3_omni/
+  #         - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
+  #         - vllm_omni/deploy/qwen3_omni_moe.yaml
+  #       timeout_in_minutes: 50
+  #       commands:
+  #         - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+  #       mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_minicpmo_4_5.py
-          - tests/e2e/online_serving/test_minicpmo_4_5.py
-          - vllm_omni/deploy/minicpmo_4_5.yaml
         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
-
-      - label: "TTS · VoxCPM2 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_voxcpm2_tts.py
-          - tests/e2e/online_serving/test_voxcpm2_tts.py
-          - vllm_omni/model_executor/models/voxcpm2/
-          - vllm_omni/deploy/voxcpm2.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_voxcpm2_tts.py tests/e2e/online_serving/test_voxcpm2_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        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
-          - vllm_omni/diffusion/models/qwen_image/
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        mirror_hardwares: h100_1
-      - label: "Diffusion · Bagel Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_bagel.py
-          - tests/e2e/online_serving/test_bagel.py
-          - vllm_omni/model_executor/models/bagel/
-          - vllm_omni/diffusion/models/bagel/
-          - vllm_omni/model_executor/stage_input_processors/bagel.py
-          - vllm_omni/deploy/bagel.yaml
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_bagel.py tests/e2e/online_serving/test_bagel.py -m 'advanced_model and cuda' --run-level 'advanced_model'
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · Qwen Image Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image.py
-          - vllm_omni/diffusion/models/qwen_image/
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        timeout_in_minutes: 50
         commands:
-          - |
-            timeout 40m bash -c '
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            '
+          - 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_1
 
-      - label: "Diffusion · Qwen Image Layered Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_layered.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: h100_1
+      # - label: "TTS · VoxCPM2 Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/offline_inference/test_voxcpm2_tts.py
+      #     - tests/e2e/online_serving/test_voxcpm2_tts.py
+      #     - vllm_omni/model_executor/models/voxcpm2/
+      #     - vllm_omni/deploy/voxcpm2.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_voxcpm2_tts.py tests/e2e/online_serving/test_voxcpm2_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+      #       "
+      #   mirror_hardwares: l4_1
 
-      - label: "Diffusion · Wan22 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_wan22_t2v.py
-          - tests/e2e/online_serving/test_wan22_t2v.py
-          - vllm_omni/diffusion/models/wan2_2/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/offline_inference/test_wan22_t2v.py tests/e2e/online_serving/test_wan22_t2v.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-            "
-        mirror_hardwares: h100_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 · Z Image Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_z_image.py
-          - tests/e2e/online_serving/test_z_image.py
-          - vllm_omni/diffusion/models/z_image/
-        timeout_in_minutes: 30
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_z_image.py tests/e2e/online_serving/test_z_image.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-        mirror_hardwares: l4_1
-      - label: "Diffusion · FLUX.2 Klein Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_flux2_klein.py
-          - vllm_omni/diffusion/models/flux2_klein/
-        timeout_in_minutes: 30
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --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
+      #     - vllm_omni/diffusion/models/qwen_image/
+      #   timeout_in_minutes: 20
+      #   commands:
+      #     - pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+      #   mirror_hardwares: h100_1
+      # - label: "Diffusion · Bagel Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/offline_inference/test_bagel.py
+      #     - tests/e2e/online_serving/test_bagel.py
+      #     - vllm_omni/model_executor/models/bagel/
+      #     - vllm_omni/diffusion/models/bagel/
+      #     - vllm_omni/model_executor/stage_input_processors/bagel.py
+      #     - vllm_omni/deploy/bagel.yaml
+      #   timeout_in_minutes: 20
+      #   commands:
+      #     - pytest -s -v tests/e2e/offline_inference/test_bagel.py tests/e2e/online_serving/test_bagel.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Qwen Image Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen_image.py
+      #     - vllm_omni/diffusion/models/qwen_image/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c '
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+      #       '
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Qwen Image Layered Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen_image_layered.py
+      #     - vllm_omni/diffusion/models/qwen_image/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Wan22 Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/offline_inference/test_wan22_t2v.py
+      #     - tests/e2e/online_serving/test_wan22_t2v.py
+      #     - vllm_omni/diffusion/models/wan2_2/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/offline_inference/test_wan22_t2v.py tests/e2e/online_serving/test_wan22_t2v.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Z Image Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/offline_inference/test_z_image.py
+      #     - tests/e2e/online_serving/test_z_image.py
+      #     - vllm_omni/diffusion/models/z_image/
+      #   timeout_in_minutes: 30
+      #   commands:
+      #     - pytest -s -v tests/e2e/offline_inference/test_z_image.py tests/e2e/online_serving/test_z_image.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+      #   mirror_hardwares: l4_1
+      # - label: "Diffusion · FLUX.2 Klein Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/offline_inference/test_flux2_klein.py
+      #     - vllm_omni/diffusion/models/flux2_klein/
+      #   timeout_in_minutes: 30
+      #   commands:
+      #     - pytest -s -v tests/e2e/offline_inference/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+      #   mirror_hardwares: l4_1

.buildkite/cuda/test-ready.yml

@@ -18,339 +18,302 @@ steps:
           - "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"
         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
+  # - 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:
+  # - group: ":card_index_dividers: Diffusion Test"
+  #   depends_on: upload-ready-pipeline
+  #   steps:
 
-      - label: "Diffusion · Offloader Test"
-        commands:
-          - timeout 40m pytest -s -v tests/diffusion/offloader -m "core_model and cuda" --run-level "core_model"
-        mirror_hardwares: l4_1
+  #     - label: "Diffusion · Offloader Test"
+  #       commands:
+  #         - timeout 40m pytest -s -v tests/diffusion/offloader -m "core_model and cuda" --run-level "core_model"
+  #       mirror_hardwares: l4_1
 
-      - label: "Diffusion · Cache Test"
-        commands:
-          - timeout 15m pytest -sv tests/diffusion/cache -m 'core_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_1
+  #     - label: "Diffusion · Cache Test"
+  #       commands:
+  #         - timeout 15m pytest -sv tests/diffusion/cache -m 'core_model and cuda' --run-level "core_model"
+  #       mirror_hardwares: l4_1
 
-      - label: "Diffusion · Batch Test"
-        commands:
-          - timeout 15m pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_1
+  #     - label: "Diffusion · Batch Test"
+  #       commands:
+  #         - timeout 15m pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
+  #       mirror_hardwares: l4_1
 
-      - label: "Diffusion · Distributed Test"
-        source_file_dependencies:
-          - tests/diffusion/distributed/
-          - vllm_omni/diffusion/distributed/
-          - vllm_omni/diffusion/attention/parallel/
-          - vllm_omni/diffusion/data.py
-          - vllm_omni/diffusion/models/
-        commands:
-          - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_4
+  #     - label: "Diffusion · Distributed Test"
+  #       source_file_dependencies:
+  #         - tests/diffusion/distributed/
+  #         - vllm_omni/diffusion/distributed/
+  #         - vllm_omni/diffusion/attention/parallel/
+  #         - vllm_omni/diffusion/data.py
+  #         - vllm_omni/diffusion/models/
+  #       commands:
+  #         - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda' --run-level "core_model"
+  #       mirror_hardwares: l4_4
 
-      - label: "Diffusion · Model Test"
-        commands:
-          - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_1
+  #     - label: "Diffusion · Model Test"
+  #       commands:
+  #         - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"
+  #       mirror_hardwares: l4_1
 
-      - label: "Diffusion · Other Test"
-        commands:
-          - timeout 15m pytest -sv tests/diffusion/*.py -m 'core_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_1
+  #     - label: "Diffusion · Other Test"
+  #       commands:
+  #         - timeout 15m pytest -sv tests/diffusion/*.py -m 'core_model and cuda' --run-level "core_model"
+  #       mirror_hardwares: l4_1
 
-  - group: ":card_index_dividers: Tiny Model Tests [Single GPU]"
-    depends_on: upload-ready-pipeline
-    steps:
+  # - group: ":card_index_dividers: Tiny Model Tests [Single GPU]"
+  #   depends_on: upload-ready-pipeline
+  #   steps:
 
-      - label: "Tiny Model Tests · Diffusion (base)"
-        commands:
-          - timeout 15m pytest -sv tests/model_tests/diffusion -m 'core_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_1
+  #     - label: "Tiny Model Tests · Diffusion (base)"
+  #       commands:
+  #         - timeout 15m pytest -sv tests/model_tests/diffusion -m 'core_model and cuda' --run-level "core_model"
+  #       mirror_hardwares: l4_1
 
-  - label: "Engine&Model Executor Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 15m bash -c "
-                  pytest -sv tests/engine/ tests/model_executor/ -m 'core_model and cuda' --run-level 'core_model'
-        "
-    mirror_hardwares: l4_1
+  # - label: "Engine&Model Executor Test"
+  #   depends_on: upload-ready-pipeline
+  #   commands:
+  #     - |
+  #       timeout 15m bash -c "
+  #                 pytest -sv tests/engine/ tests/model_executor/ -m 'core_model and cuda' --run-level 'core_model'
+  #       "
+  #   mirror_hardwares: l4_1
 
-  - label: "Distributed Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 20m bash -c "
-          pytest -sv tests/distributed/ -m 'core_model and cuda and L4' --run-level 'core_model'
-        "
-    mirror_hardwares: l4_1
+  # - label: "Distributed Test"
+  #   depends_on: upload-ready-pipeline
+  #   commands:
+  #     - |
+  #       timeout 20m bash -c "
+  #         pytest -sv tests/distributed/ -m 'core_model and cuda and L4' --run-level 'core_model'
+  #       "
+  #   mirror_hardwares: l4_1
 
-  - label: "Entrypoints Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 30m bash -c "
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -sv tests/entrypoints/ -m 'core_model and cuda' --run-level 'core_model'
-        "
-    mirror_hardwares: l4_1
+  # - label: "Entrypoints Test"
+  #   depends_on: upload-ready-pipeline
+  #   commands:
+  #     - |
+  #       timeout 30m bash -c "
+  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  #         pytest -sv tests/entrypoints/ -m 'core_model and cuda' --run-level 'core_model'
+  #       "
+  #   mirror_hardwares: l4_1
 
   - group: ":card_index_dividers: E2E Test"
     depends_on: upload-ready-pipeline
     steps:
 
-      - 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/
-          - vllm_omni/diffusion/
-          - vllm_omni/engine/
-          - vllm_omni/entrypoints/
-          - vllm_omni/inputs/
-          - vllm_omni/outputs/
-        commands:
-          - |
-            timeout 20m bash -c "
-              pytest -s -v tests/e2e/offline_inference/rlhf_test/test_verl_omni_e2e.py
-            "
-        mirror_hardwares: l4_1
+      # - 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/
+      #     - vllm_omni/diffusion/
+      #     - vllm_omni/engine/
+      #     - vllm_omni/entrypoints/
+      #     - vllm_omni/inputs/
+      #     - vllm_omni/outputs/
+      #   commands:
+      #     - |
+      #       timeout 20m bash -c "
+      #         pytest -s -v tests/e2e/offline_inference/rlhf_test/test_verl_omni_e2e.py
+      #       "
+      #   mirror_hardwares: l4_1
 
-      - label: "Omni · Qwen3-Omni Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_omni.py
-          - vllm_omni/model_executor/models/qwen3_omni/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
-          - vllm_omni/deploy/qwen3_omni_moe.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_2
+      # - label: "Omni · Qwen3-Omni Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen3_omni.py
+      #     - vllm_omni/model_executor/models/qwen3_omni/
+      #     - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
+      #     - vllm_omni/deploy/qwen3_omni_moe.yaml
+      #   commands:
+      #     - |
+      #       timeout 20m bash -c "
+      #         pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Offline Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_minicpmo_4_5.py
-          - vllm_omni/deploy/minicpmo_4_5.yaml
         commands:
           - |
             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_1
 
       - label: "Omni · MiniCPM-o 4.5 Online Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_minicpmo_4_5.py
-          - vllm_omni/deploy/minicpmo_4_5.yaml
         commands:
           - |
             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_1
 
-      - label: "TTS · VoxCPM2 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_voxcpm2_tts.py
-          - vllm_omni/model_executor/models/voxcpm2/
-          - vllm_omni/deploy/voxcpm2.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/online_serving/test_voxcpm2_tts.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "TTS · Qwen3-TTS CustomVoice Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
         commands:
           - |
             timeout 20m bash -c "
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "TTS · Qwen3-TTS Base Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_base.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              export VLLM_OMNI_USE_V2_RUNNER=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "TTS · Higgs-Audio-V3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v3.py
-          - vllm_omni/model_executor/models/higgs_audio_v3/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
-          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "Diffusion · Qwen Image Edit Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_edit.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 20m bash -c "
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'core_model' --run-level 'core_model'
+              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_1
 
-      - label: "Diffusion · Bagel Test"
+      - label: "Omni · MiniCPM-o 4.5 Perf Test"
+        timeout_in_minutes: 60
         source_file_dependencies:
-          - tests/e2e/online_serving/test_bagel.py
-          - vllm_omni/model_executor/models/bagel/
-          - vllm_omni/diffusion/models/bagel/
-          - vllm_omni/model_executor/stage_input_processors/bagel.py
-          - vllm_omni/deploy/bagel.yaml
+          - tests/dfx/perf/tests/test_minicpmo_4_5_ready.json
+          - vllm_omni/deploy/minicpmo_4_5.yaml
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/models/step_audio2/
+          - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+          - vllm_omni/model_executor/models/cosyvoice3/utils.py
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
         commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_bagel.py -m 'core_model' --run-level 'core_model'
-            "
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_ready.json -m 'core_model and cuda'
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · Wan22 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_wan22_t2v.py
-          - vllm_omni/diffusion/models/wan2_2/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
+      # - label: "TTS · VoxCPM2 Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_voxcpm2_tts.py
+      #     - vllm_omni/model_executor/models/voxcpm2/
+      #     - vllm_omni/deploy/voxcpm2.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/online_serving/test_voxcpm2_tts.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: l4_1
 
-      - label: "Diffusion · Cosmos3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_cosmos3.py
-          - vllm_omni/diffusion/models/cosmos3/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_cosmos3.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
+      # - label: "TTS · Qwen3-TTS CustomVoice Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
+      #     - vllm_omni/model_executor/models/qwen3_tts/
+      #     - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+      #     - vllm_omni/deploy/qwen3_tts.yaml
+      #   commands:
+      #     - |
+      #       timeout 20m bash -c "
+      #         export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+      #         pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: l4_1
 
-      - label: "Diffusion · Qwen Image Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
+      # - label: "TTS · Qwen3-TTS Base Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen3_tts_base.py
+      #     - vllm_omni/model_executor/models/qwen3_tts/
+      #     - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+      #     - vllm_omni/deploy/qwen3_tts.yaml
+      #   commands:
+      #     - |
+      #       timeout 20m bash -c "
+      #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      #         export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+      #         export VLLM_OMNI_USE_V2_RUNNER=1
+      #         pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: l4_1
 
-      - label: "Diffusion · Qwen Image Layered Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_layered.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
+      # - label: "TTS · Higgs-Audio-V3 Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_higgs_audio_v3.py
+      #     - vllm_omni/model_executor/models/higgs_audio_v3/
+      #     - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
+      #     - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
+      #   commands:
+      #     - |
+      #       timeout 20m bash -c "
+      #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      #         export VLLM_USE_DEEP_GEMM=0
+      #         export VLLM_MOE_USE_DEEP_GEMM=0
+      #         pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: l4_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
+      # - label: "Diffusion · Qwen Image Edit Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen_image_edit.py
+      #     - vllm_omni/diffusion/models/qwen_image/
+      #   commands:
+      #     - |
+      #       timeout 20m bash -c "
+      #         pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Bagel Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_bagel.py
+      #     - vllm_omni/model_executor/models/bagel/
+      #     - vllm_omni/diffusion/models/bagel/
+      #     - vllm_omni/model_executor/stage_input_processors/bagel.py
+      #     - vllm_omni/deploy/bagel.yaml
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_bagel.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Wan22 Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_wan22_t2v.py
+      #     - vllm_omni/diffusion/models/wan2_2/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Cosmos3 Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_cosmos3.py
+      #     - vllm_omni/diffusion/models/cosmos3/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_cosmos3.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Qwen Image Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen_image.py
+      #     - vllm_omni/diffusion/models/qwen_image/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   mirror_hardwares: h100_1
+
+      # - label: "Diffusion · Qwen Image Layered Test"
+      #   source_file_dependencies:
+      #     - tests/e2e/online_serving/test_qwen_image_layered.py
+      #     - vllm_omni/diffusion/models/qwen_image/
+      #   commands:
+      #     - |
+      #       timeout 40m bash -c "
+      #         export VLLM_IMAGE_FETCH_TIMEOUT=60
+      #         pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_model'
+      #       "
+      #   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/scripts/bootstrap-intel-omni.sh

@@ -4,6 +4,10 @@
 
 set -euo pipefail
 
+# minicpm-challenge branch: skip entire Intel/XPU CI (no pipeline upload => no image build / tests).
+echo "--- Skipping Intel/XPU CI on this branch (including image build)"
+exit 0
+
 source .buildkite/common/scripts/resolve_skip_ci.sh
 
 upload_pipeline() {

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

@@ -1,25 +1,118 @@
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+  HF_HUB_DOWNLOAD_TIMEOUT: 300
+  HF_HUB_ETAG_TIMEOUT: 60
+
 steps:
-  - group: ":card_index_dividers: Diffusion X2V Model Test"
-    key: nightly-diffusion-x2v-group
+  # - group: ":card_index_dividers: Diffusion X2V Model Test"
+  #   key: nightly-diffusion-x2v-group
+  #   depends_on: upload-nightly-pipeline
+  #   if: >-
+  #     build.env("NIGHTLY") == "1" ||
+  #     build.pull_request.labels includes "nightly-test" ||
+  #     build.pull_request.labels includes "diffusion-x2v-test"
+  #   steps:
+  #     - label: ":full_moon: Diffusion X2V · Function Test"
+  #       mirror_hardwares: a2b3_npu_4
+  #       env:
+  #         VLLM_WORKER_MULTIPROC_METHOD: spawn
+  #         HF_TOKEN: "${HF_TOKEN}"
+  #       commands:
+  #         - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k npu_i2v --run-level "full_model"
+  # - group: ":card_index_dividers: TTS Model Test"
+  #   key: nightly-tts-test-group
+  #   depends_on: upload-nightly-pipeline
+  #   if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
+  #   steps:
+  #     - label: ":full_moon: TTS · Perf Test"
+  #       mirror_hardwares: a3_npu_2
+  #       env:
+  #         VLLM_WORKER_MULTIPROC_METHOD: spawn
+  #         VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+  #         HF_TOKEN: "${HF_TOKEN}"
+  #       commands:
+  #         - export BENCHMARK_DIR=tests/dfx/perf/results
+  #         - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+  #         - |
+  #           set +e
+  #           pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m npu --test-config-file tests/dfx/perf/tests/test_tts.json
+  #           EXIT=$$?
+  #           buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+  #           exit $$EXIT
+
+  - group: ":card_index_dividers: Omni Model Test"
+    key: nightly-omni-group
     depends_on: upload-nightly-pipeline
-    if: >-
-      build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2v-test"
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "omni-test"
     steps:
-      - label: ":full_moon: Diffusion X2V · Function Test"
-        mirror_hardwares: a2b3_npu_4
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 Test"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
           HF_TOKEN: "${HF_TOKEN}"
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k npu_i2v --run-level "full_model"
-  - group: ":card_index_dividers: TTS Model Test"
-    key: nightly-tts-test-group
-    depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
-    steps:
-      - label: ":full_moon: TTS · Perf Test"
+          - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5_expansion.py -m 'full_model and npu and A3 and omni' --run-level 'full_model'
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Duplex Test"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          # NPU has no merge-level pipeline, so advanced_model duplex coverage
+          # rides on nightly alongside full_model.
+          #
+          # test_duplex_soft_interrupt is CUDA-only and carries no npu mark, so
+          # the marker filter below drops it: it needs real-time duplex
+          # throughput that the current NPU stack does not reach.
+          - >-
+            pytest -s -v
+            tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+            tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+            -m "(full_model or advanced_model) and npu and A3 and omni"
+            --run-level full_model
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Accuracy Test"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - export SEED_TTS_WER_EVAL=1
+          - export SEED_TTS_SIM_EVAL=1
+          - export SEED_TTS_EVAL_DEVICE=npu:1
+          - |
+            set +e
+            pytest -s -v tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py -k "not videomme" -m "full_model and npu and A3" --run-level full_model
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/e2e/accuracy/minicpmo_4_5/results/*.json"
+            exit $$EXIT
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Video-MME Accuracy Test"
+        key: nightly-omni-accuracy-minicpmo-4-5-videomme-npu
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          # Video-MME needs no repo override here: the NPU runners pre-stage the dataset under
+          # the lmms-eval mirror, which is now VIDEOMME_DEFAULT_HF_REPO, so snapshot_download
+          # resolves it straight out of the hub cache.
+          - |
+            set +e
+            pytest -s -v tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py -k videomme -m "full_model and npu and A3" --run-level full_model
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/e2e/accuracy/minicpmo_4_5/results/*.json"
+            exit $$EXIT
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test"
+        key: nightly-omni-performance-minicpmo-4-5-npu
+        timeout_in_minutes: 180
         mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
@@ -27,32 +120,42 @@ steps:
           HF_TOKEN: "${HF_TOKEN}"
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
-          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m npu --test-config-file tests/dfx/perf/tests/test_tts.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "npu and A3" --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5.json
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-
-  - group: ":card_index_dividers: Omni Model Test"
-    key: nightly-omni-group
-    depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "omni-test"
-    steps:
-      - label: ":robot: Omni · MiniCPM-o 4.5 Test"
-        agents:
-          queue: "ascend-a3"
-          resource_class: "npu-2"
-        image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a3"
-        plugins:
-          - kubernetes:
-              podSpecPatch:
-                imagePullSecrets:
-                  - name: swr-secret
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test (A2)"
+        key: nightly-omni-performance-minicpmo-4-5-npu-a2
+        timeout_in_minutes: 180
+        mirror_hardwares: a2b3_npu_1
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+          HF_HUB_OFFLINE: "1"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "npu and A2" --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Duplex Seed-TTS Perf Test"
+        key: nightly-omni-performance-minicpmo-4-5-duplex-seed-tts-npu
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
           VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
           HF_TOKEN: "${HF_TOKEN}"
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5_expansion.py -m 'full_model' --run-level 'full_model'
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "npu and A3" --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT

.buildkite/npu/test-npu-ready.yml

@@ -3,32 +3,41 @@ steps:
     key: ready-omni-group
     depends_on: upload-ready-pipeline
     steps:
-      - label: ":robot: Omni · MiniCPM-o 4.5 Test"
-        agents:
-          queue: "ascend-a3"
-          resource_class: "npu-2"
-        image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a3"
-        plugins:
-          - kubernetes:
-              podSpecPatch:
-                imagePullSecrets:
-                  - name: swr-secret
+      - label: "Omni · MiniCPM-o 4.5 Test"
+        mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
           VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
           HF_TOKEN: "${HF_TOKEN}"
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py -m 'core_model' --run-level 'core_model'
+          - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py -m 'core_model and npu and A3 and omni' --run-level 'core_model'
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        mirror_hardwares: a3_npu_2
+        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_minicpmo_4_5_duplex.py -m 'core_model and npu and A3 and omni' --run-level 'core_model'
+      - label: "Omni · MiniCPM-o 4.5 Perf Test"
+        timeout_in_minutes: 60
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_ready.json -m 'core_model and npu and A3 and omni'
 
-  - group: ":card_index_dividers: TTS Model Test"
-    key: ready-tts-group
-    depends_on: upload-ready-pipeline
-    steps:
-      - label: ":sound: TTS · Qwen3-TTS CustomVoice Test"
-        mirror_hardwares: a2b3_npu_1
-        env:
-          VLLM_WORKER_MULTIPROC_METHOD: spawn
-          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
-          HF_TOKEN: "${HF_TOKEN}"
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'
+  # - group: ":card_index_dividers: TTS Model Test"
+  #   key: ready-tts-group
+  #   depends_on: upload-ready-pipeline
+  #   steps:
+  #     - label: ":sound: TTS · Qwen3-TTS CustomVoice Test"
+  #       mirror_hardwares: a2b3_npu_1
+  #       env:
+  #         VLLM_WORKER_MULTIPROC_METHOD: spawn
+  #         VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+  #         HF_TOKEN: "${HF_TOKEN}"
+  #       commands:
+  #         - pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'

docker/Dockerfile.npu.ci

@@ -32,6 +32,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "pytest-cov>=4.0.0" \
       "pytest-mock>=3.10.0" \
       "openpyxl>=3.0.0" \
+      "opencc==1.4.1" \
       "psutil>=7.2.0" \
       "soundfile>=0.13.1" \
       "pyttsx3>=2.99" \
@@ -40,6 +41,11 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "hyperpyyaml>=1.2.0" \
       "onnx>=1.16.0" \
       "librosa>=0.10.0" \
+      "jiwer>=4.0.0" \
+      "zhon>=2.1.1" \
+      "zhconv>=1.4.3" \
+      "funasr>=1.3.0" \
+      "datasets>=3.0.0" \
       --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
       --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
       --default-timeout=60 && \

docker/Dockerfile.npu.ci.a3

@@ -32,6 +32,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "pytest-cov>=4.0.0" \
       "pytest-mock>=3.10.0" \
       "openpyxl>=3.0.0" \
+      "opencc==1.4.1" \
       "psutil>=7.2.0" \
       "soundfile>=0.13.1" \
       "pyttsx3>=2.99" \
@@ -40,6 +41,11 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "hyperpyyaml>=1.2.0" \
       "onnx>=1.16.0" \
       "librosa>=0.10.0" \
+      "jiwer>=4.0.0" \
+      "zhon>=2.1.1" \
+      "zhconv>=1.4.3" \
+      "funasr>=1.3.0" \
+      "datasets>=3.0.0" \
       --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/contributing/ci/test_examples/l4_performance_tests.inc.md

@@ -44,9 +44,11 @@ Pass **`--test-config-file`** to run one JSON file as-is, or omit it for the bul
             "ignore_eos": true,
             "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
             "baseline": {
-                "mean_ttft_ms": [500, 800],
-                "mean_audio_ttfp_ms": [2000, 3500],
-                "mean_audio_rtf": [0.25, 0.35]
+                "H100": {
+                    "mean_ttft_ms": [500, 800],
+                    "mean_audio_ttfp_ms": [2000, 3500],
+                    "mean_audio_rtf": [0.25, 0.35]
+                }
             }
         }
     ]
@@ -142,10 +144,6 @@ pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py \
   --test-config-file tests/dfx/perf/tests/test_bagel_vllm_omni.json
 pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
   --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
-
-# Optional baseline assertion (default off)
-pytest -sv tests/dfx/perf/scripts/run_diffusion_benchmark.py --assert-baseline \
-  --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
 ```
 
 See also [Markers for Tests](#markers-for-tests) for registered hardware markers (`H100`, `L4`, `cuda`, `distributed_cuda`, …).
@@ -202,4 +200,4 @@ You can add any benchmark running parameters you need here. For all optional par
 | num_prompts     | int / array | Yes      | 10,[10, 20, 30] | Number of requests. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of qps or max_concurrency, e.g., [10,10,10]. If an array is used, its length must match the number of qps or max_concurrency. |
 | request_rate    | float / array | No  | 0.5, [0.5, 1, inf] | Queries per second. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of num_prompts, e.g., [1,1,1]. If an array is used, its length must match the number of num_prompts.                          |
 | max_concurrency | int / array | No       | 1, [1, 2, 3]    | Maximum concurrent in-flight requests. Same array / expansion rules as `request_rate` (mutually exclusive with QPS mode).                                                                                                                                                                                                             |
-| baseline        | object      | No       | see above       | Optional per-metric thresholds; keys must match benchmark output fields. Scalar, list (per sweep step), or object (keyed by concurrency or QPS string).  
+| baseline        | object      | No       | see below       | Optional thresholds. **Hardware-nested only**: top-level keys must be `[hardware-resource]` pytest markers from `pyproject.toml` (`get_hardware_mark_list()`). Under each hardware key, metric names are free-form; values are scalars or sweep-aligned lists. Example: `{"H100": {"throughput_qps": […], "custom_stage_ms": 1.0}, "A3": {…}}`. Runtime filename aliases (`_RUNTIME_DEVICE_ALIASES`) are separate and do not expand this allowlist. |

examples/offline_inference/minicpmo/README.md

@@ -1,8 +1,9 @@
 # MiniCPM-o 4.5: Offline inference
 
-Two-stage pipeline: **thinker** (multimodal understanding) → **talker + Token2Wav**
+Three-stage pipeline: **Thinker** → **Talker** (codec) → **Code2Wav**
 (24 kHz speech). Deploy config auto-loads from
-`vllm_omni/deploy/minicpmo_4_5.yaml` (2-GPU default).
+`vllm_omni/deploy/minicpmo_4_5.yaml` (single-GPU default). For higher
+throughput, pass `--deploy-config vllm_omni/deploy/minicpmo_4_5_2gpu.yaml`.
 
 ## Setup
 
@@ -72,7 +73,7 @@ Supported `--query-type` values:
 
 ```bash
 python end2end.py --query-type text \
-    --deploy-config /path/to/vllm_omni/deploy/minicpmo_4_5_8x4090.yaml
+    --deploy-config /path/to/vllm_omni/deploy/minicpmo_4_5_2gpu.yaml
 ```
 
 ## Notes
@@ -83,7 +84,8 @@ python end2end.py --query-type text \
 - Output WAV is **24 kHz mono**.
 - Placeholders in the prompt are MiniCPM-style:
   `(<image>./</image>)`, `(<audio>./</audio>)`, `(<video>./</video>)`.
-- Default layout needs **2 GPUs**. Async chunking is off in the bundled YAMLs.
+- Default layout needs **1 GPU**. Use `minicpmo_4_5_2gpu.yaml` for the
+  recommended two-GPU continuous-batching layout.
 
 ## Online serving
 

examples/offline_inference/minicpmo/end2end.py

@@ -16,7 +16,6 @@ import numpy as np
 import soundfile as sf
 import vllm
 from PIL import Image
-from vllm import SamplingParams
 from vllm.assets.audio import AudioAsset
 from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset, video_to_ndarrays
@@ -26,8 +25,6 @@ from vllm.multimodal.media.audio import load_audio
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.utils.tracking_parser import TrackingArgumentParser
 
-SEED = 42
-
 default_system = (
     "You are MiniCPM-o, a helpful multimodal assistant that can "
     "understand images, audio and video, and respond in text and speech."
@@ -285,27 +282,17 @@ def main(args):
     omni_kwargs["trust_remote_code"] = True
     omni = Omni(**omni_kwargs)
 
-    # Stage 0 (thinker): multimodal understanding → text (+ TTS span when enabled).
-    thinker_sampling_params = SamplingParams(
-        temperature=0.0,
-        top_p=1.0,
-        top_k=-1,
-        max_tokens=2048,
-        seed=SEED,
-        detokenize=True,
-        repetition_penalty=1.1,
-    )
-    # Stage 1 (talker + Token2Wav): max_tokens=1 satisfies the scheduler;
-    # waveform is produced in-process by Token2Wav.
-    talker_sampling_params = SamplingParams(
-        temperature=0.0,
-        top_p=1.0,
-        top_k=-1,
-        max_tokens=1,
-        seed=SEED,
-        detokenize=False,
-    )
-    sampling_params_list = [thinker_sampling_params, talker_sampling_params][: omni.num_stages]
+    # Per-stage sampling defaults live in the deploy YAML
+    # (``vllm_omni/deploy/minicpmo_4_5*.yaml``, ``default_sampling_params``),
+    # which is the single source of truth that tracks the pipeline topology.
+    # Passing None makes ``resolve_sampling_params_list`` load exactly those,
+    # so adding or removing a stage cannot desynchronize this example.
+    #
+    # This used to hardcode a two-element list built for the old two-stage
+    # topology (thinker → talker+Token2Wav). Since Code2Wav became its own
+    # Stage 2, that list was both the wrong length and wrong in content
+    # (talker ``max_tokens=1`` truncated the codec stream to one frame).
+    sampling_params_list = None
 
     if args.txt_prompts is None:
         prompts = [query_result.inputs for _ in range(args.num_prompts)]

examples/online_serving/minicpmo/README.md

@@ -23,17 +23,17 @@ 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
-GPU 0 (90%) and colocates the Talker (55%) and Code2Wav (35%) on GPU 1. That
-profile admits at most four concurrent sequences per stage.
+logical device 0 with memory budgets of 55%, 15%, and 18%, leaving headroom
+for runtime kernels such as the HiFi-GAN vocoder's cuDNN workspace. The
+profile admits at most four sequences per stage. 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 +89,18 @@ 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/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()

pyproject.toml

@@ -243,13 +243,16 @@ markers = [
     "npu: Tests that run on NPU/Ascend (auto-added)",
     "musa: Tests that run on MUSA/Moore Threads (auto-added)",
     # specified computation resources marks (auto-added)
-    "H100: Tests that require H100 GPU",
-    "L4: Tests that require L4 GPU",
-    "MI325: Tests that require MI325 GPU (AMD/ROCm)",
-    "B60: Tests that require Intel Arc Pro B60 XPU",
-    "S5000: Tests that require S5000 GPU (Moore Threads/MUSA)",
-    "A2: Tests that require A2 NPU",
-    "A3: Tests that require A3 NPU",
+    # Tag with [hardware-resource] for tests.helpers.mark.get_hardware_mark_list().
+    "H100: [hardware-resource] Tests that require H100 GPU",
+    "H200: [hardware-resource] Tests that require H200 GPU",
+    "L4: [hardware-resource] Tests that require L4 GPU",
+    "B200: [hardware-resource] Tests that require B200 GPU",
+    "MI325: [hardware-resource] Tests that require MI325 GPU (AMD/ROCm)",
+    "B60: [hardware-resource] Tests that require Intel Arc Pro B60 XPU",
+    "S5000: [hardware-resource] Tests that require S5000 GPU (Moore Threads/MUSA)",
+    "A2: [hardware-resource] Tests that require A2 NPU",
+    "A3: [hardware-resource] Tests that require A3 NPU",
     "distributed_cuda: Tests that require multi cards on CUDA platform",
     "distributed_rocm: Tests that require multi cards on ROCm platform",
     "distributed_xpu: Tests that require multi cards on XPU platform",

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

@@ -31,9 +31,8 @@ also provided.
   - 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_2gpu.yaml`](../../vllm_omni/deploy/minicpmo_4_5_2gpu.yaml)
+  - 3-GPU layout:
     [`vllm_omni/deploy/minicpmo_4_5_3gpu.yaml`](../../vllm_omni/deploy/minicpmo_4_5_3gpu.yaml)
   - 8x RTX 4090 layout:
     [`vllm_omni/deploy/minicpmo_4_5_8x4090.yaml`](../../vllm_omni/deploy/minicpmo_4_5_8x4090.yaml)
@@ -68,8 +67,8 @@ 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
-two-GPU profile. The removed fused two-stage implementation is not retained as
+single-GPU entry point; `minicpmo_4_5_2gpu.yaml` is the recommended
+two-GPU continuous-batching profile. The removed fused two-stage implementation is not retained as
 a fallback because it would duplicate state machines and correctness paths.
 
 ## GPU
@@ -79,9 +78,9 @@ 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
-up to four sequences. Startup video profiling is bounded to 32 frames per
-video. Use the two-GPU profile for production throughput.
+`gpu_memory_utilization` budgets are 0.55, 0.15, and 0.18 so concurrent
+Code2Wav activations keep residual headroom. All stages admit up to four
+sequences. Use the two-GPU profile for production throughput.
 
 #### Environment
 
@@ -105,7 +104,7 @@ 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 +174,18 @@ 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)**:
 
@@ -202,8 +203,8 @@ 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;
+- Memory budget: Thinker, Talker, and Code2Wav reserve 0.55, 0.15, and
+  0.18 of GPU 0 so concurrent Code2Wav activations keep residual headroom;
   all three model processes still share one CUDA device.
 - `--trust-remote-code` is required — the HF repo ships a custom
   `MiniCPMO` config / model class.
@@ -214,14 +215,13 @@ speech output (TTS)"** checkbox on / off.
   contention. Talker AR
   state and Code2Wav caches are request-owned; Code2Wav batches only
   exact-shape-compatible chunks and does not fall back to serial decode.
-- `limit_mm_per_prompt.video.num_frames: 32` bounds startup dummy profiling,
-  not runtime media decoding. Use `media_io_kwargs.video.num_frames` when a
-  matching URL/file video sampling limit is required.
+- Stage 0 pins `interleave_mm_strings=true` and
+  `media_io_kwargs.video={fps:1,num_frames:128}` for Daily-Omni MiniCPM packs.
 - `StageRequestStats.batch_size` is a request-scoped placeholder, not the
   scheduler's execution batch.
 - Single-GPU co-location trades throughput for hardware density: Stage 0/1
   CUDA Graph replay and eager Stage 2 vocoder kernels compete across three
-  CUDA contexts. Use the 8x4090 config or a custom multi-GPU mapping for
+  CUDA contexts. Use `minicpmo_4_5_2gpu.yaml` or a custom multi-GPU mapping for
   throughput-sensitive serving.
 
 ### 8 x RTX 4090 24GB (consumer-GPU layout)

recipes/Tencent/HunyuanImage-3.0-Instruct.md

@@ -164,10 +164,11 @@ curl -s http://localhost:8091/v1/chat/completions \
 PR [#2495](https://github.com/vllm-project/vllm-omni/pull/2495) adds
 performance CI configs for the same DiT-only settings. The CI step is
 currently opt-in (gated by `RUN_HUNYUAN_IMAGE3_PERF=1`) with `soft_fail`
-enabled, intended for initial data collection. Performance assertions are
-skipped (`skip-performance-assertion: true`); the baseline values in the
-JSON configs are reference-only and will be promoted to regression gates
-once enough nightly data has been collected.
+enabled, intended for initial data collection. Perf runners no longer
+compare metrics against JSON baselines at assert time; the `baseline`
+blocks in the configs are reference-only and are copied into result JSON
+for reporting. Regression gates can be reintroduced once enough nightly
+data has been collected.
 
 The user-facing equivalent is to launch one of the CLI commands above and
 generate 1024x1024 images with 50 denoising steps.

tests/benchmarks/patch/test_patch.py

@@ -6,13 +6,21 @@ Unit tests for patch.py
 """
 
 import asyncio
+import base64
 import json
+import time
+from types import SimpleNamespace
 
 import pytest
 from pytest_mock import MockerFixture
 from vllm.benchmarks.lib.endpoint_request_func import RequestFuncInput
 
-from vllm_omni.benchmarks.patch.patch import MixRequestFuncOutput, async_request_openai_chat_omni_completions
+from vllm_omni.benchmarks.patch.patch import (
+    MixRequestFuncOutput,
+    async_request_openai_chat_omni_completions,
+    async_request_openai_realtime_duplex,
+)
+from vllm_omni.experimental.fullduplex.client import RealtimeEventCollector
 
 pytestmark = [pytest.mark.core_model, pytest.mark.benchmark, pytest.mark.cpu]
 
@@ -40,6 +48,156 @@ class MockResponse:
         pass
 
 
+@pytest.mark.asyncio
+async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch):
+    class FakeRealtimeClient:
+        last_instance = None
+
+        def __init__(self, url):
+            assert url == "ws://localhost:8000/v1/realtime?duplex=1"
+            self.events = RealtimeEventCollector()
+            self.configure_kwargs = None
+            self.sent = []
+            self.response_count = 0
+            self.ack_count = 0
+            type(self).last_instance = self
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *_args):
+            return None
+
+        async def configure(self, model, **kwargs):
+            assert model == "openbmb/MiniCPM-o-4_5"
+            self.configure_kwargs = kwargs
+
+        async def send(self, event):
+            self.sent.append(event)
+            if event["type"] != "response.create":
+                return
+            self.response_count += 1
+            response_id = f"resp-{self.response_count}"
+            now = time.monotonic()
+            self.events.add(
+                {"type": "response.created", "response": {"id": response_id}},
+                received_at_s=now + 0.01,
+            )
+            self.events.add(
+                {
+                    "type": "response.audio_transcript.delta",
+                    "response_id": response_id,
+                    "delta": f"turn {self.response_count}",
+                },
+                received_at_s=now + 0.02,
+            )
+            self.events.add(
+                {
+                    "type": "response.audio.delta",
+                    "response_id": response_id,
+                    "delta": base64.b64encode(b"\x00\x00" * 2400).decode(),
+                    "sample_rate_hz": 24_000,
+                    "metadata": {"audio_duration_ms": 100},
+                },
+                received_at_s=now + 0.03,
+            )
+            self.events.add(
+                {"type": "response.done", "response": {"id": response_id}},
+                received_at_s=now + 0.04,
+            )
+
+        async def acknowledge_playback(self):
+            self.ack_count += 1
+
+        async def close_session(self, **_kwargs):
+            return None
+
+    monkeypatch.setattr(
+        "vllm_omni.benchmarks.patch.patch.RealtimeDuplexClient",
+        FakeRealtimeClient,
+    )
+    request_input = RequestFuncInput(
+        model="openbmb/MiniCPM-o-4_5",
+        model_name="openbmb/MiniCPM-o-4_5",
+        prompt="hello",
+        api_url="http://localhost:8000/v1/realtime",
+        prompt_len=1,
+        output_len=20,
+        logprobs=None,
+        multi_modal_content=None,
+        ignore_eos=False,
+        extra_body={"save_duplex_request_metrics": True},
+    )
+    request_input.seed_tts_speech_extra = {"ref_audio": "data:audio/wav;base64,AAAA"}
+    request_input.seed_tts_system_prompt = "Speak exactly."
+    request_input.seed_tts_turns = tuple(
+        SimpleNamespace(utterance_id=f"utt-{index}", target_text=f"text {index}") for index in range(4)
+    )
+
+    output = await async_request_openai_realtime_duplex(
+        request_input,
+        session=None,
+    )
+
+    client = FakeRealtimeClient.last_instance
+    assert client.configure_kwargs["native_duplex"] is False
+    assert client.configure_kwargs["extra_body"] == {
+        "ref_audio": "data:audio/wav;base64,AAAA",
+    }
+    assert client.sent == [
+        event
+        for index in range(4)
+        for event in (
+            {
+                "type": "conversation.item.create",
+                "item": {
+                    "type": "message",
+                    "role": "user",
+                    "content": [{"type": "input_text", "text": f"text {index}"}],
+                },
+            },
+            {"type": "response.create"},
+        )
+    ]
+    assert client.ack_count == 4
+    assert output.success is True
+    assert output.generated_text == "turn 1 turn 2 turn 3 turn 4"
+    assert output.audio_duration == pytest.approx(0.4)
+    assert output.ttft > 0
+    assert output.audio_ttfp > output.ttft
+    assert output.audio_rtf > 0
+    assert output.latency > 0
+    assert output.tts_turn_pcm_bytes == [b"\x00\x00" * 2400] * 4
+    assert output.tts_output_pcm_bytes == b"\x00\x00" * 9600
+    session_id = output.duplex_request_metrics[0]["session_id"]
+    assert len(output.duplex_request_metrics) == 4
+    for index, metrics in enumerate(output.duplex_request_metrics):
+        assert metrics == {
+            "session_id": session_id,
+            "request_index": index,
+            "utterance_id": f"utt-{index}",
+            "response_id": f"resp-{index + 1}",
+            "source": "client_monotonic_receive",
+            "measurement_origin": {
+                "ttft": "conversation.item.create client send to first non-empty text delta",
+                "ttfp": "conversation.item.create client send to first audio packet",
+                "rtf": "request-start-to-last-audio receive time divided by emitted audio duration",
+            },
+            "ttft_ms": pytest.approx(20.0, abs=2.0),
+            "ttfp_ms": pytest.approx(30.0, abs=2.0),
+            "rtf": pytest.approx(0.3, abs=0.03),
+            "audio_generation_ms": pytest.approx(30.0, abs=2.0),
+            "audio_duration_ms": 100.0,
+        }
+    assert output.duplex_session_metrics == {
+        "session_id": session_id,
+        "audio_turn_count": 4,
+        "mean_ttft_ms": pytest.approx(20.0, abs=2.0),
+        "mean_ttfp_ms": pytest.approx(30.0, abs=2.0),
+        "mean_rtf": pytest.approx(0.3, abs=0.03),
+    }
+
+
 def create_sse_chunk(data_dict):
     """Helper to create SSE formatted chunk"""
     return f"data: {json.dumps(data_dict)}\n\n".encode()

tests/benchmarks/test_accuracy_bench_utils.py

@@ -42,7 +42,10 @@ from benchmarks.accuracy.text_to_image.gbench import (
     summarize_generated_records as summarize_gebench_generated_records,
     summarize_gebench_results,
 )
-from tests.e2e.accuracy.qwen3_omni.qwen3_omni_acc_bench_core import seed_tts_bench_argv
+from tests.e2e.accuracy.qwen3_omni.qwen3_omni_acc_bench_core import (
+    build_serve_common_argv,
+    seed_tts_bench_argv,
+)
 from tests.e2e.accuracy.qwen3_omni.run_qwen_omni_acc_benchmark import sync_dataset_env_from_ns
 from vllm_omni.benchmarks.data_modules.seed_tts_dataset import resolve_seed_tts_root
 
@@ -57,6 +60,25 @@ def test_seed_tts_bench_argv_preserves_hf_repo_id_from_env(monkeypatch):
     assert argv[dataset_idx + 1] == "zhaochenyang20/seed-tts-eval"
 
 
+def test_build_serve_common_argv_accepts_realtime_seed_tts_route(tmp_path):
+    argv = build_serve_common_argv(
+        host="127.0.0.1",
+        port=8000,
+        model="openbmb/MiniCPM-o-4_5",
+        num_prompts=4,
+        max_concurrency=1,
+        num_warmups=0,
+        percentile_metrics="ttft,audio_ttfp,audio_rtf",
+        result_dir=tmp_path,
+        result_filename="result.json",
+        backend="openai-realtime-tts",
+        endpoint="/v1/realtime",
+    )
+
+    assert argv[argv.index("--backend") + 1] == "openai-realtime-tts"
+    assert argv[argv.index("--endpoint") + 1] == "/v1/realtime"
+
+
 def test_sync_dataset_env_preserves_seed_tts_hf_repo_id(monkeypatch):
     ns = argparse.Namespace(
         daily_omni_repo=None,

tests/benchmarks/test_daily_omni_local_source.py

@@ -0,0 +1,152 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Offline Daily-Omni sources: local mirrors must not fall back to the Hub id."""
+
+import json
+import os
+import tarfile
+from pathlib import Path
+
+import pytest
+
+from vllm_omni.benchmarks.data_modules.daily_omni_dataset import (
+    DailyOmniDataset,
+    daily_omni_local_qa_json,
+    daily_omni_local_videos_dir,
+    ensure_daily_omni_hub_root,
+    resolve_daily_omni_local_root,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.benchmark]
+
+_HUB_ID = "liarliar/Daily-Omni"
+
+
+def _write_qa_json(root: Path) -> Path:
+    qa = root / "qa.json"
+    qa.write_text(json.dumps([{"Question": "q", "Answer": "A", "video_id": "vid0"}]), encoding="utf-8")
+    return qa
+
+
+def _stage_offline_hub_cache(tmp_path: Path, monkeypatch: pytest.MonkeyPatch, *, with_qa: bool = True) -> Path:
+    """Build a hub cache holding ``liarliar/Daily-Omni`` and cut off all network access.
+
+    Reproduces the air-gapped NPU runner: the dataset is pre-staged with ``hf download``, so it
+    only exists in the *hub* cache -- never in the ``datasets`` cache that ``load_dataset`` reads.
+    """
+    constants = pytest.importorskip("huggingface_hub").constants
+
+    cache = tmp_path / "hub"
+    repo_dir = cache / "datasets--liarliar--Daily-Omni"
+    snapshot = repo_dir / "snapshots" / "deadbeef"
+    snapshot.mkdir(parents=True)
+    if with_qa:
+        _write_qa_json(snapshot)
+    (repo_dir / "refs").mkdir()
+    (repo_dir / "refs" / "main").write_text("deadbeef", encoding="utf-8")
+
+    monkeypatch.setattr(constants, "HF_HUB_CACHE", str(cache))
+    monkeypatch.setattr(constants, "HF_HUB_OFFLINE", True)
+    return snapshot
+
+
+def test_resolve_local_root_returns_none_for_hub_id() -> None:
+    assert resolve_daily_omni_local_root("liarliar/Daily-Omni") is None
+    assert resolve_daily_omni_local_root("") is None
+    assert resolve_daily_omni_local_root(None) is None
+
+
+def test_plain_local_mirror(tmp_path: Path) -> None:
+    _write_qa_json(tmp_path)
+    (tmp_path / "Videos" / "vid0").mkdir(parents=True)
+
+    root = resolve_daily_omni_local_root(str(tmp_path))
+    assert root == tmp_path.resolve()
+    assert daily_omni_local_qa_json(root) == (tmp_path / "qa.json").resolve()
+    assert daily_omni_local_videos_dir(root) == root / "Videos"
+
+
+def test_hf_cache_dir_resolves_to_snapshot(tmp_path: Path) -> None:
+    cache_dir = tmp_path / "datasets--liarliar--Daily-Omni"
+    snapshot = cache_dir / "snapshots" / "deadbeef"
+    snapshot.mkdir(parents=True)
+    _write_qa_json(snapshot)
+    (cache_dir / "refs").mkdir()
+    (cache_dir / "refs" / "main").write_text("deadbeef", encoding="utf-8")
+
+    root = resolve_daily_omni_local_root(str(cache_dir))
+    assert root == snapshot.resolve()
+    assert daily_omni_local_qa_json(root) == (snapshot / "qa.json").resolve()
+
+
+def test_hf_cache_dir_without_refs_picks_newest_snapshot(tmp_path: Path) -> None:
+    cache_dir = tmp_path / "datasets--liarliar--Daily-Omni"
+    older = cache_dir / "snapshots" / "aaa"
+    newer = cache_dir / "snapshots" / "bbb"
+    older.mkdir(parents=True)
+    newer.mkdir(parents=True)
+    os.utime(older, (1, 1))
+    os.utime(newer, (10, 10))
+
+    assert resolve_daily_omni_local_root(str(cache_dir)) == newer.resolve()
+
+
+def test_videos_tar_is_extracted(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("HF_HOME", str(tmp_path / "hf_home"))
+    root = tmp_path / "mirror"
+    (root / "Videos" / "vid0").mkdir(parents=True)
+    (root / "Videos" / "vid0" / "vid0_video.mp4").write_bytes(b"fake")
+    with tarfile.open(root / "Videos.tar", "w") as tf:
+        tf.add(root / "Videos", arcname="Videos")
+    # Only the tarball is left in the mirror.
+    for child in (root / "Videos").rglob("*"):
+        if child.is_file():
+            child.unlink()
+    (root / "Videos" / "vid0").rmdir()
+    (root / "Videos").rmdir()
+
+    videos_dir = daily_omni_local_videos_dir(root)
+    assert videos_dir is not None
+    assert (videos_dir / "vid0" / "vid0_video.mp4").is_file()
+
+    # Second call reuses the extraction instead of redoing it.
+    assert daily_omni_local_videos_dir(root) == videos_dir
+
+
+def test_missing_qa_json_reports_none(tmp_path: Path) -> None:
+    root = resolve_daily_omni_local_root(str(tmp_path))
+    assert daily_omni_local_qa_json(root) is None
+    assert daily_omni_local_videos_dir(root) is None
+
+
+def test_hub_id_resolves_to_offline_cache(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    snapshot = _stage_offline_hub_cache(tmp_path, monkeypatch)
+
+    root = ensure_daily_omni_hub_root(_HUB_ID)
+
+    assert root == snapshot.resolve()
+    assert daily_omni_local_qa_json(root) == (snapshot / "qa.json").resolve()
+
+
+def test_dataset_loads_hub_id_without_network(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    """The regression this file exists for: a Hub id must not require ``load_dataset``."""
+    snapshot = _stage_offline_hub_cache(tmp_path, monkeypatch)
+    # Any use of `datasets` here means we went back to the path that cannot read the hub cache.
+    monkeypatch.setattr("vllm_omni.benchmarks.data_modules.daily_omni_dataset.load_dataset", None)
+
+    dataset = DailyOmniDataset(dataset_path=_HUB_ID, random_seed=0)
+
+    assert dataset.qa_json_path == (snapshot / "qa.json").resolve()
+    assert [row["video_id"] for row in dataset.data] == ["vid0"]
+
+
+def test_uncached_hub_id_reports_both_attempts(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    _stage_offline_hub_cache(tmp_path, monkeypatch, with_qa=False)
+    monkeypatch.setattr("vllm_omni.benchmarks.data_modules.daily_omni_dataset.load_dataset", None)
+
+    with pytest.raises(RuntimeError) as excinfo:
+        DailyOmniDataset(dataset_path="someone-else/Daily-Omni", random_seed=0)
+
+    message = str(excinfo.value)
+    assert "someone-else/Daily-Omni" in message
+    assert "snapshot_download failed" in message
+    assert "--daily-omni-qa-json" in message

tests/benchmarks/test_daily_omni_pack_modes.py

@@ -0,0 +1,237 @@
+"""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 no 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), (7, 7), (3, 5), (1, 1)])
+def test_uniform_sample_indices_matches_numpy_linspace(n: int, k: int) -> None:
+    got = _uniform_sample_indices(n, k)
+    if n <= k:
+        assert got == list(range(n))
+    else:
+        assert got == [int(i) for i in np.linspace(0, n - 1, k, dtype=int)]
+
+
+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_omits_qwen_system_message(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)
+
+    messages = ds._build_daily_omni_openai_messages(parts, "q?", {"A": "a"})
+
+    assert [m["role"] for m in messages] == ["user"]
+
+
+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/benchmarks/test_seed_tts_dataset_variants.py

@@ -28,10 +28,13 @@ if _MODULE_NAME not in sys.modules:
     _spec.loader.exec_module(_mod)
 
 from vllm_omni.benchmarks.data_modules.seed_tts_dataset import (  # noqa: E402
+    SeedTTSDataset,
     SeedTTSDesignDataset,
     SeedTTSDesignSampleRequest,
+    SeedTTSSampleRequest,
     SeedTTSTextDataset,
     SeedTTSTextSampleRequest,
+    SeedTTSTurn,
 )
 
 # ---------------------------------------------------------------------------
@@ -70,6 +73,94 @@ def mock_tokenizer(mocker):
 # ---------------------------------------------------------------------------
 
 
+def test_seed_tts_dataset_groups_four_turns_with_one_reference(seed_tts_root, mock_tokenizer):
+    ds = SeedTTSDataset(
+        dataset_path=str(seed_tts_root),
+        random_seed=0,
+        locale="en",
+        disable_shuffle=True,
+    )
+
+    requests = ds.sample(
+        mock_tokenizer,
+        num_requests=1,
+        turns_per_session=4,
+        no_oversample=True,
+    )
+
+    assert len(requests) == 1
+    request = requests[0]
+    assert [turn.utterance_id for turn in request.seed_tts_turns] == [
+        "utt000",
+        "utt001",
+        "utt002",
+        "utt003",
+    ]
+    assert [turn.target_text for turn in request.seed_tts_turns] == [
+        "target text 0",
+        "target text 1",
+        "target text 2",
+        "target text 3",
+    ]
+    assert request.seed_tts_speech_extra["ref_text"] == "ref text 0"
+    assert request.seed_tts_ref_wav_path.endswith("utt000.wav")
+
+
+def test_seed_tts_eval_expands_grouped_pcm_by_turn():
+    from vllm_omni.benchmarks.data_modules.seed_tts_eval import (
+        _expand_seed_tts_turn_outputs,
+    )
+
+    request = SeedTTSSampleRequest(
+        prompt="target 0",
+        prompt_len=4,
+        expected_output_len=100,
+        multi_modal_data=None,
+        request_id="session-0",
+        seed_tts_turns=tuple(
+            SeedTTSTurn(utterance_id=f"utt-{index}", target_text=f"target {index}") for index in range(4)
+        ),
+    )
+    output = types.SimpleNamespace(
+        success=True,
+        tts_turn_pcm_bytes=[bytes([index]) for index in range(4)],
+    )
+
+    requests, outputs = _expand_seed_tts_turn_outputs([request], [output])
+
+    assert [item.prompt for item in requests] == [f"target {index}" for index in range(4)]
+    assert [item.seed_tts_utterance_id for item in requests] == [f"utt-{index}" for index in range(4)]
+    assert [item.tts_output_pcm_bytes for item in outputs] == [bytes([index]) for index in range(4)]
+
+
+def test_seed_tts_eval_keeps_session_pcm_for_single_turn_chat():
+    """Chat-omni Seed-TTS sets session PCM only; expand must not wipe it."""
+    from vllm_omni.benchmarks.data_modules.seed_tts_eval import (
+        _expand_seed_tts_turn_outputs,
+    )
+
+    request = SeedTTSSampleRequest(
+        prompt="target 0",
+        prompt_len=4,
+        expected_output_len=100,
+        multi_modal_data=None,
+        request_id="session-0",
+        seed_tts_turns=(SeedTTSTurn(utterance_id="utt-0", target_text="target 0"),),
+    )
+    pcm = b"\x01\x02\x03\x04"
+    output = types.SimpleNamespace(
+        success=True,
+        tts_output_pcm_bytes=pcm,
+        tts_turn_pcm_bytes=None,
+    )
+
+    requests, outputs = _expand_seed_tts_turn_outputs([request], [output])
+
+    assert len(requests) == 1
+    assert requests[0].prompt == "target 0"
+    assert outputs[0].tts_output_pcm_bytes == pcm
+
+
 def test_seed_tts_text_dataset_omits_ref_audio(seed_tts_root, mock_tokenizer):
     ds = SeedTTSTextDataset(
         dataset_path=str(seed_tts_root),

tests/core/sched/test_omni_ar_scheduler_streaming.py

@@ -242,9 +242,52 @@ 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

@@ -54,6 +54,18 @@ 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(
@@ -133,6 +145,21 @@ 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()

tests/dfx/conftest.py

@@ -3,15 +3,13 @@ import os
 import re
 import subprocess
 import threading
-import time
 from datetime import datetime
 from pathlib import Path
 from typing import Any
 
 import pytest
 
-from tests.dfx.reliability.helpers import list_remote_process_pids_by_pattern, post_chat_completions_raw
-from tests.helpers.mark import hardware_marks
+from tests.helpers.mark import get_hardware_mark_list, hardware_marks
 from tests.helpers.runtime import OmniServerParams
 from tests.helpers.stage_config import modify_stage_config
 from vllm_omni.platforms import current_omni_platform
@@ -119,32 +117,6 @@ def create_unique_server_pytest_params(
     ]
 
 
-def create_benchmark_pytest_params(
-    benchmark_configs: list[dict[str, Any]],
-    server_to_benchmark_mapping: dict[str, dict],
-) -> list[Any]:
-    """Like :func:`create_benchmark_indices`, but wrap each index in ``pytest.param`` with JSON marks."""
-    marks_by_name = _marks_by_test_name(benchmark_configs)
-    params: list[Any] = []
-    seen: set[str] = set()
-    for config in benchmark_configs:
-        test_name = config["test_name"]
-        if test_name in seen:
-            continue
-        seen.add(test_name)
-        params_list = get_benchmark_params_for_server(test_name, server_to_benchmark_mapping)
-        id_suffixes = _unique_benchmark_param_id_suffixes(params_list)
-        for idx, id_suffix in enumerate(id_suffixes):
-            params.append(
-                pytest.param(
-                    (test_name, idx),
-                    marks=marks_by_name.get(test_name, []),
-                    id=f"{test_name}-{id_suffix}",
-                )
-            )
-    return params
-
-
 def create_paired_benchmark_pytest_params(
     server_entries: list[tuple[Any, str]],
     params_by_test_name: dict[str, list[Any]],
@@ -357,10 +329,6 @@ def supports_video_generation(model_name: str) -> bool:
     return any(key in lower for key in ("wan", "video", "i2v", "t2v"))
 
 
-def supports_chat_generation(model_name: str) -> bool:
-    return not supports_video_generation(model_name)
-
-
 def parse_stage_devices(stage_config_path: str) -> str:
     text = Path(stage_config_path).read_text(encoding="utf-8")
     raw_devices: list[str] = re.findall(r"^\s*devices:\s*\"?([0-9,\s]+)\"?\s*$", text, flags=re.MULTILINE)
@@ -389,48 +357,6 @@ def assert_fault_exception(exc: Exception, error_keywords: tuple[str, ...]) -> N
     assert any(key in text for key in error_keywords), f"unexpected error under fault injection: {exc}"
 
 
-def wait_chat_request_ready(host: str, port: int, model: str, timeout_sec: int = 180) -> None:
-    """Poll a minimal chat request until success."""
-    deadline = time.time() + timeout_sec
-    payload = json.dumps(
-        {
-            "model": model,
-            "messages": [{"role": "user", "content": "Say hello in one short sentence."}],
-            "stream": False,
-            "modalities": ["text"],
-        }
-    )
-    last_error: str | None = None
-    while time.time() < deadline:
-        try:
-            status, body = post_chat_completions_raw(host, port, payload)
-            if status == 200:
-                return
-            last_error = f"http={status} body={body[:200]!r}"
-        except Exception as exc:  # noqa: BLE001
-            last_error = str(exc)
-        time.sleep(2)
-    raise TimeoutError(f"runtime-teardown warmup request did not succeed within {timeout_sec}s: {last_error}")
-
-
-def assert_no_extra_worker_processes(
-    baseline_pids: set[int],
-    worker_pattern: str,
-    timeout_sec: int = 60,
-) -> None:
-    """Ensure no extra worker PIDs remain after teardown."""
-    deadline = time.time() + timeout_sec
-    while time.time() < deadline:
-        current = set(list_remote_process_pids_by_pattern(worker_pattern))
-        extra = current - baseline_pids
-        if not extra:
-            return
-        time.sleep(2)
-    current = set(list_remote_process_pids_by_pattern(worker_pattern))
-    extra = sorted(current - baseline_pids)
-    assert not extra, f"orphan worker processes remain after container teardown: {extra}"
-
-
 def create_test_parameter_mapping(configs: list[dict[str, Any]]) -> dict[str, dict]:
     mapping = {}
     for config in configs:
@@ -520,36 +446,16 @@ def _unique_benchmark_param_id_suffixes(params_list: list[dict[str, Any]]) -> li
     return unique
 
 
-def extract_mark_resource_label(mark_field: Any) -> str:
-    """Return a filename-safe hardware label from ``mark.hardware_marks.res`` values.
-
-    Example: ``{"cuda": "H100"}`` -> ``"H100"``; multiple platforms join with ``-``.
-
-    Prefer :func:`get_runtime_resource_label` for perf result filenames so labels
-    reflect the machine that actually ran the benchmark.
-    """
-    if isinstance(mark_field, list):
-        for item in mark_field:
-            if isinstance(item, dict) and "hardware_marks" in item:
-                return extract_mark_resource_label(item)
-        return "na"
-    if not isinstance(mark_field, dict):
-        return "na"
-    hw = mark_field.get("hardware_marks")
-    if not isinstance(hw, dict):
-        return "na"
-    res = hw.get("res")
-    if not isinstance(res, dict) or not res:
-        return "na"
-    labels = [_safe_filename_token(value) for value in res.values()]
-    return "-".join(labels) if labels else "na"
-
-
-_KNOWN_RUNTIME_RESOURCE_TOKENS: tuple[str, ...] = (
+# Device marketing-name tokens for ``get_runtime_resource_label`` only.
+# Longer / more specific tokens must appear before shorter substrings
+# (e.g. ``H200`` before ``H20``, ``910B4`` before ``910``).
+_RUNTIME_DEVICE_ALIASES: tuple[str, ...] = (
     "H100",
     "H800",
     "H200",
     "H20",
+    "B200",
+    "GB200",
     "L40S",
     "L40",
     "L4",
@@ -578,7 +484,7 @@ def _normalize_runtime_device_label(raw: str) -> str:
     if not raw or not str(raw).strip():
         return "na"
     upper = str(raw).upper()
-    for token in _KNOWN_RUNTIME_RESOURCE_TOKENS:
+    for token in _RUNTIME_DEVICE_ALIASES:
         if token.upper() in upper:
             return _safe_filename_token(token)
     compact = re.sub(r"[^a-zA-Z0-9]+", "", str(raw))
@@ -631,34 +537,41 @@ def resource_label_for_filename(resource_label: str | None) -> str:
     return token
 
 
-def extract_configs_resource_label(configs: list[dict[str, Any]]) -> str:
-    """Return runtime hardware label for perf result filenames."""
-    del configs
-    return get_runtime_resource_label()
+def is_hardware_nested_baseline(baseline: dict[str, Any]) -> bool:
+    """True for ``{"H100": {"metric": ...}, "A3": {...}}`` baselines.
+
+    Detection is structural + hardware-label allowlist (not metric-name hints):
+
+    - every top-level value is a non-empty dict (per-hardware metric map);
+    - every top-level key is in :func:`tests.helpers.mark.get_hardware_mark_list`
+      (``[hardware-resource]`` markers from ``pyproject.toml``).
+
+    Metric names under each hardware bucket are unconstrained.
+    Runtime device aliases (:data:`_RUNTIME_DEVICE_ALIASES`) are unrelated.
+    """
+    if not baseline:
+        return False
+    if not all(isinstance(value, dict) and bool(value) for value in baseline.values()):
+        return False
+    allowed = {label.upper() for label in get_hardware_mark_list()}
+    return all((token := str(key).strip()) and token.upper() in allowed for key in baseline.keys())
 
 
 def resolve_baseline_value(
     baseline_raw: Any,
     *,
     sweep_index: int | None,
-    max_concurrency: Any = None,
-    request_rate: Any = None,
 ) -> Any:
-    """Pick the baseline threshold for this sweep step."""
+    """Pick the baseline threshold for one metric at this sweep step.
+
+    Accepts a sweep-aligned list/tuple (indexed by ``sweep_index``) or a scalar.
+    """
     if baseline_raw is None:
-        return 100000
+        return None
     if isinstance(baseline_raw, dict):
-        if max_concurrency is not None:
-            for key in (max_concurrency, str(max_concurrency)):
-                if key in baseline_raw:
-                    return baseline_raw[key]
-        if request_rate is not None:
-            for key in (request_rate, str(request_rate)):
-                if key in baseline_raw:
-                    return baseline_raw[key]
-        raise KeyError(
-            f"baseline dict has no key for max_concurrency={max_concurrency!r} "
-            f"or request_rate={request_rate!r}; keys={list(baseline_raw.keys())!r}"
+        raise TypeError(
+            "per-metric baseline dict keyed by concurrency/request-rate is not supported; "
+            "use a sweep-aligned list or a scalar under each hardware bucket"
         )
     if isinstance(baseline_raw, (list, tuple)):
         if sweep_index is None:
@@ -669,22 +582,62 @@ def resolve_baseline_value(
     return baseline_raw
 
 
-def _baseline_thresholds_for_step(
-    baseline_data: dict[str, Any],
+def resolve_baseline_for_sweep(
+    baseline_config: dict[str, Any] | None,
     *,
     sweep_index: int | None = None,
-    max_concurrency: Any = None,
-    request_rate: Any = None,
 ) -> dict[str, Any]:
-    """Resolve baseline config to one threshold per metric for this iteration."""
-    return {
-        metric_name: resolve_baseline_value(
-            baseline_raw,
-            sweep_index=sweep_index,
-            max_concurrency=max_concurrency,
-            request_rate=request_rate,
+    """Resolve sweep-dependent baselines while keeping every hardware bucket.
+
+    Expects hardware-nested input only::
+
+        {"H100": {"throughput_qps": [a, b, c]}, "A3": {"throughput_qps": [x, y, z]}}
+
+    with ``sweep_index=2`` (e.g. concurrency 32) becomes::
+
+        {"H100": {"throughput_qps": c}, "A3": {"throughput_qps": z}}
+
+    Per-metric values may be a sweep-aligned list or a scalar.
+    """
+    if not baseline_config:
+        return {}
+    if not is_hardware_nested_baseline(baseline_config):
+        allowed = sorted(get_hardware_mark_list())
+        unknown = [
+            str(key).strip()
+            for key in baseline_config.keys()
+            if str(key).strip().upper() not in {label.upper() for label in allowed}
+        ]
+        nested_shape = all(isinstance(value, dict) and bool(value) for value in baseline_config.values())
+        if nested_shape and unknown:
+            raise ValueError(
+                "baseline top-level keys must be [hardware-resource] pytest markers from "
+                "pyproject.toml (tool.pytest.ini_options.markers). "
+                f"Unknown hardware label(s): {unknown!r}. "
+                f"Known markers: {allowed!r}. "
+                "Add a marker such as '\"A100: [hardware-resource] Tests that require A100 GPU\"' "
+                "to pyproject.toml, then retry."
+            )
+        raise ValueError(
+            "baseline must be hardware-nested with [hardware-resource] marker names as "
+            "top-level keys, e.g. "
+            '{"H100": {"throughput_qps": [...]}, "A3": {"custom_metric": 1.0}}. '
+            f"got top-level keys={list(baseline_config.keys())!r}; "
+            f"known markers={allowed!r}. "
+            "If you need a new hardware bucket, add it to pyproject.toml markers with "
+            "the [hardware-resource] tag."
         )
-        for metric_name, baseline_raw in baseline_data.items()
+
+    return {
+        hardware: {
+            metric_name: resolve_baseline_value(
+                baseline_raw,
+                sweep_index=sweep_index,
+            )
+            for metric_name, baseline_raw in metrics.items()
+        }
+        for hardware, metrics in baseline_config.items()
+        if isinstance(metrics, dict)
     }
 
 
@@ -697,17 +650,17 @@ def run_benchmark(
     *,
     baseline_config: dict[str, Any] | None = None,
     sweep_index: int | None = None,
-    request_rate: Any | None = None,
-    max_concurrency: Any | None = None,
     random_input_len: Any | None = None,
     random_output_len: Any | None = None,
     resource_label: str | None = None,
 ) -> dict[str, Any]:
     """Run one ``vllm bench serve --omni`` iteration and return parsed metrics.
 
-    After ``vllm bench`` writes the JSON, ``result["baseline"]`` holds the resolved per-metric thresholds
-    (when ``baseline_config`` is provided). If the benchmark exits without writing a result file,
-    ``result_omni_template.json`` is used as a fallback.
+    After ``vllm bench`` writes the JSON, ``result["baseline"]`` stores the
+    sweep-resolved baseline (when ``baseline_config`` is provided). Hardware-nested
+    maps keep every hardware bucket, but each metric is reduced to the value for
+    this concurrency / request-rate step. If the benchmark exits without writing a
+    result file, ``result_omni_template.json`` is used as a fallback.
     """
     current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
     ri = _safe_filename_token(random_input_len)
@@ -774,11 +727,10 @@ def run_benchmark(
             result = json.load(f)
 
     if baseline_config:
-        result["baseline"] = _baseline_thresholds_for_step(
+        # Keep every hardware bucket; resolve list metrics to this sweep step.
+        result["baseline"] = resolve_baseline_for_sweep(
             baseline_config,
             sweep_index=sweep_index,
-            request_rate=request_rate,
-            max_concurrency=max_concurrency,
         )
     else:
         result["baseline"] = {}
@@ -800,12 +752,3 @@ def pytest_addoption(parser: pytest.Parser) -> None:
         default=None,
         help=("Path to benchmark config JSON. Example: --test-config-file tests/dfx/perf/tests/test_tts.json"),
     )
-    parser.addoption(
-        "--assert-baseline",
-        action="store_true",
-        default=False,
-        help=(
-            "When set, omni/diffusion perf runners compare metrics against the baseline block in the JSON config "
-            "(default: off)."
-        ),
-    )

tests/dfx/perf/scripts/diffusion_result_template.json

@@ -21,10 +21,12 @@
       "num-input-images": 0,
       "enable-negative-prompt": false,
       "baseline": {
-        "throughput_qps": 0,
-        "latency_mean": 0,
-        "peak_memory_mb_max": 0,
-        "peak_memory_mb_mean": 0
+        "H100": {
+          "throughput_qps": 0,
+          "latency_mean": 0,
+          "peak_memory_mb_max": 0,
+          "peak_memory_mb_mean": 0
+        }
       }
     },
     "result": {

tests/dfx/perf/scripts/result_omni_template.json

@@ -46,9 +46,11 @@
   "median_audio_duration_s": 0,
   "p99_audio_duration_s": 0,
   "baseline": {
-    "mean_ttft_ms": 0,
-    "mean_audio_ttfp_ms": 0,
-    "mean_audio_rtf": 0
+    "H100": {
+      "mean_ttft_ms": 0,
+      "mean_audio_ttfp_ms": 0,
+      "mean_audio_rtf": 0
+    }
   },
   "random_input_len": 0,
   "random_output_len": 0

tests/dfx/perf/scripts/run_benchmark.py

@@ -1,6 +1,8 @@
 import json
 import os
 import threading
+from collections.abc import Callable
+from contextlib import AbstractContextManager, ExitStack, contextmanager
 from pathlib import Path
 from typing import Any
 
@@ -13,15 +15,10 @@ from tests.dfx.conftest import (
     get_runtime_resource_label,
     is_diffusion_perf_config,
     load_benchmark_configs,
-    resolve_baseline_value,
     run_benchmark,
 )
 from tests.helpers.runtime import OmniServer
 
-# Compare metrics to each test JSON ``baseline`` block only when pytest is run with ``--assert-baseline``
-# (registered in ``tests/dfx/conftest.py``; default: off). ``run_benchmark`` and ``_resolve_baseline_value`` are
-# defined in the same module.
-#
 # Optional JSON field ``mark`` is applied as pytest marks via
 # ``create_paired_omni_benchmark_pytest_params`` (e.g. ``"mark": [{"hardware_marks":
 # {"res": {"cuda": "H100"}, "num_cards": 2}}, "full_model", "omni"]``).
@@ -63,35 +60,78 @@ paired_benchmark_params = create_paired_omni_benchmark_pytest_params(BENCHMARK_C
 _omni_server_lock = threading.Lock()
 
 
+class _SingleActiveContext:
+    """Reuse one active context while its configuration key is unchanged."""
+
+    def __init__(self) -> None:
+        self._key: Any = None
+        self._stack: ExitStack | None = None
+        self._value: Any = None
+
+    def acquire(self, key: Any, factory: Callable[[], AbstractContextManager[Any]]) -> Any:
+        if self._stack is not None and key == self._key:
+            return self._value
+
+        self.close()
+        stack = ExitStack()
+        value = stack.enter_context(factory())
+        self._key = key
+        self._stack = stack
+        self._value = value
+        return value
+
+    def close(self) -> None:
+        stack = self._stack
+        self._key = None
+        self._stack = None
+        self._value = None
+        if stack is not None:
+            stack.close()
+
+
+@contextmanager
+def _start_omni_server(server_param):
+    test_name, model, stage_config_path, stage_overrides, extra_cli_args, use_omni = server_param
+
+    print(f"Starting OmniServer with test: {test_name}, model: {model}")
+
+    server_args: list[str] = []
+    if use_omni:
+        server_args += ["--stage-init-timeout", "600", "--init-timeout", "900"]
+    # --deploy-config and --stage-overrides compose at the CLI (see vllm_omni/entrypoints/utils.py):
+    # deploy-config sets the base; stage-overrides are applied on top. Both can be set.
+    if stage_config_path:
+        server_args = ["--deploy-config", stage_config_path] + server_args
+    if stage_overrides:
+        server_args = ["--stage-overrides", stage_overrides] + server_args
+    if extra_cli_args:
+        server_args = list(extra_cli_args) + server_args
+    with OmniServer(model, server_args, use_omni=use_omni) as server:
+        server.test_name = test_name
+        print("OmniServer started successfully")
+        yield server
+        print("OmniServer stopping...")
+
+    print("OmniServer stopped")
+
+
 @pytest.fixture(scope="module")
-def omni_server(request):
+def omni_server_context():
     """Start vLLM-Omni server as a subprocess with actual model weights.
-    Uses session scope so the server starts only once for the entire test session.
+    Reuse it for adjacent benchmark cases with the same server configuration.
     Multi-stage initialization can take 10-20+ minutes.
     """
     with _omni_server_lock:
-        test_name, model, stage_config_path, stage_overrides, extra_cli_args, use_omni = request.param
+        active_context = _SingleActiveContext()
+        try:
+            yield active_context
+        finally:
+            active_context.close()
 
-        print(f"Starting OmniServer with test: {test_name}, model: {model}")
 
-        server_args: list[str] = []
-        if use_omni:
-            server_args += ["--stage-init-timeout", "600", "--init-timeout", "900"]
-        # --deploy-config and --stage-overrides compose at the CLI (see vllm_omni/entrypoints/utils.py):
-        # deploy-config sets the base; stage-overrides are applied on top. Both can be set.
-        if stage_config_path:
-            server_args = ["--deploy-config", stage_config_path] + server_args
-        if stage_overrides:
-            server_args = ["--stage-overrides", stage_overrides] + server_args
-        if extra_cli_args:
-            server_args = list(extra_cli_args) + server_args
-        with OmniServer(model, server_args, use_omni=use_omni) as server:
-            server.test_name = test_name
-            print("OmniServer started successfully")
-            yield server
-            print("OmniServer stopping...")
-
-        print("OmniServer stopped")
+@pytest.fixture
+def omni_server(request, omni_server_context):
+    return omni_server_context.acquire(request.param, lambda: _start_omni_server(request.param))
 
 
 @pytest.fixture
@@ -117,36 +157,19 @@ def benchmark_params(request):
     }
 
 
-def assert_result(
-    result,
-    params,
-    num_prompt,
-    *,
-    assert_baseline: bool,
-    sweep_index: int | None = None,
-    max_concurrency: Any | None = None,
-    request_rate: Any | None = None,
-) -> None:
+def assert_result(result, params, num_prompt) -> None:
     assert result["completed"] == num_prompt, "Request failures exist"
-    if not assert_baseline:
-        return
-    baseline_data = params.get("baseline", {})
-    for metric_name, baseline_raw in baseline_data.items():
-        current_value = result[metric_name]
-        baseline_value = resolve_baseline_value(
-            baseline_raw,
-            sweep_index=sweep_index,
-            max_concurrency=max_concurrency,
-            request_rate=request_rate,
+    expected_audio_turns = params.get("expected_duplex_audio_turns_per_session")
+    if expected_audio_turns is not None:
+        session_metrics = result.get("duplex_session_metrics")
+        assert isinstance(session_metrics, list), "Duplex session metrics are missing"
+        assert len(session_metrics) == num_prompt, (
+            f"Expected {num_prompt} duplex session metric rows, got {len(session_metrics)}"
         )
-        if "throughput" in metric_name:
-            if current_value <= baseline_value:
-                print(
-                    f"ERROR: Throughput test results were below baseline: {metric_name}: {current_value} > {baseline_value}"
-                )
-        else:
-            if current_value >= baseline_value:
-                print(f"ERROR: Test results exceeded baseline: {metric_name}: {current_value} < {baseline_value}")
+        assert all(
+            isinstance(metric, dict) and metric.get("audio_turn_count") == expected_audio_turns
+            for metric in session_metrics
+        ), f"Not every duplex session emitted {expected_audio_turns} audio turns"
 
 
 @pytest.mark.benchmark
@@ -155,7 +178,7 @@ def assert_result(
     paired_benchmark_params,
     indirect=["omni_server", "benchmark_params"],
 )
-def test_performance_benchmark(omni_server, benchmark_params, request):
+def test_performance_benchmark(omni_server, benchmark_params):
     test_name = benchmark_params["test_name"]
     params = benchmark_params["params"]
     dataset_name = params.get("dataset_name", "")
@@ -167,7 +190,6 @@ def test_performance_benchmark(omni_server, benchmark_params, request):
     print(f"Running benchmark for model: {model}")
     print(f"Benchmark parameters: {benchmark_params}")
 
-    assert_baseline = request.config.getoption("--assert-baseline", default=False)
     resource_label = get_runtime_resource_label()
 
     def to_list(value, default=None):
@@ -201,6 +223,7 @@ def test_performance_benchmark(omni_server, benchmark_params, request):
         "enabled",
         "eval_phase",
         "trust_remote_code",
+        "expected_duplex_audio_turns_per_session",
     }
 
     for key, value in params.items():
@@ -226,7 +249,7 @@ def test_performance_benchmark(omni_server, benchmark_params, request):
         break
 
     # QPS / request-rate sweep
-    for i, (qps, num_prompt) in enumerate(zip(qps_list, num_prompt_list)):
+    for sweep_index, (qps, num_prompt) in enumerate(zip(qps_list, num_prompt_list)):
         args = args + ["--request-rate", str(qps), "--num-prompts", str(num_prompt)]
         result = run_benchmark(
             args=args,
@@ -235,24 +258,15 @@ def test_performance_benchmark(omni_server, benchmark_params, request):
             dataset_name=dataset_name,
             num_prompt=num_prompt,
             baseline_config=params.get("baseline"),
-            sweep_index=i,
-            request_rate=qps,
-            max_concurrency=None,
+            sweep_index=sweep_index,
             random_input_len=params.get("random_input_len"),
             random_output_len=params.get("random_output_len"),
             resource_label=resource_label,
         )
-        assert_result(
-            result,
-            params,
-            num_prompt,
-            assert_baseline=assert_baseline,
-            sweep_index=i,
-            request_rate=qps,
-        )
+        assert_result(result, params, num_prompt)
 
     # concurrency test
-    for i, (concurrency, num_prompt) in enumerate(zip(max_concurrency_list, num_prompt_list)):
+    for sweep_index, (concurrency, num_prompt) in enumerate(zip(max_concurrency_list, num_prompt_list)):
         args = args + ["--max-concurrency", str(concurrency), "--num-prompts", str(num_prompt), "--request-rate", "inf"]
         result = run_benchmark(
             args=args,
@@ -261,18 +275,9 @@ def test_performance_benchmark(omni_server, benchmark_params, request):
             dataset_name=dataset_name,
             num_prompt=num_prompt,
             baseline_config=params.get("baseline"),
-            sweep_index=i,
-            request_rate=None,
-            max_concurrency=concurrency,
+            sweep_index=sweep_index,
             random_input_len=params.get("random_input_len"),
             random_output_len=params.get("random_output_len"),
             resource_label=resource_label,
         )
-        assert_result(
-            result,
-            params,
-            num_prompt,
-            assert_baseline=assert_baseline,
-            sweep_index=i,
-            max_concurrency=concurrency,
-        )
+        assert_result(result, params, num_prompt)

tests/dfx/perf/scripts/run_diffusion_benchmark.py

@@ -13,8 +13,6 @@ A config JSON file may be passed via --test-config-file. If omitted, every ``*.j
   pytest run_diffusion_benchmark.py -m "diffusion"
   pytest run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
 
-Optional: ``--assert-baseline`` compares metrics to the ``baseline`` block in each benchmark entry (default: off).
-
 Optional JSON field ``mark`` is applied as pytest marks on that case via
 ``pytest.param`` (e.g. ``"mark": [{"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 1}}, "full_model", "diffusion"]``).
 
@@ -46,6 +44,7 @@ from tests.dfx.conftest import (
     get_runtime_resource_label,
     hardware_json_value,
     is_diffusion_perf_config,
+    resolve_baseline_for_sweep,
     resolve_pytest_marks,
     resource_label_for_filename,
 )
@@ -274,11 +273,6 @@ def _write_result_record(record: dict[str, Any]) -> Path:
     return target
 
 
-def _append_to_aggregated_file(record: dict[str, Any]) -> None:
-    """Backward-compatible wrapper; prefer :func:`_write_result_record`."""
-    _write_result_record(record)
-
-
 _server_lock = threading.Lock()
 
 # ---------------------------------------------------------------------------
@@ -355,6 +349,17 @@ def _kill_process_tree(pid: int) -> None:
 # ---------------------------------------------------------------------------
 
 
+def _resolve_offline_model(model: str) -> str:
+    """Under HF_HUB_OFFLINE, resolve a HF repo id to its local snapshot dir."""
+    import huggingface_hub
+
+    if not model or os.path.isdir(model) or not huggingface_hub.constants.HF_HUB_OFFLINE:
+        return model
+    from huggingface_hub import snapshot_download
+
+    return snapshot_download(model, local_files_only=True)
+
+
 class DiffusionServer:
     """Start a vLLM-Omni diffusion model server as a subprocess.
 
@@ -372,7 +377,7 @@ class DiffusionServer:
         port: int | None = None,
     ) -> None:
         self.server_cfg: dict[str, Any] = server_cfg
-        self.model = server_cfg["model"]
+        self.model = _resolve_offline_model(server_cfg["model"])
         self.serve_args = server_cfg["serve_args"]
         self.host = "127.0.0.1"
         self.port = port if port is not None else get_open_port(self.host)
@@ -667,7 +672,7 @@ def run_benchmark(
     with tempfile.NamedTemporaryFile(mode="w", suffix=".json", prefix="diffusion_bench_tmp_", delete=False) as tmp:
         tmp_result_file = Path(tmp.name)
 
-    exclude_keys = {"baseline", "dataset", "task", "name", "skip-performance-assertion"}
+    exclude_keys = {"baseline", "dataset", "task", "name"}
 
     cmd = [
         sys.executable,
@@ -761,12 +766,17 @@ def run_benchmark(
     completed = metrics.get("completed_requests", metrics.get("completed", 0))
     failed = metrics.get("failed_requests", metrics.get("failed", 0))
 
+    # Persist sweep-resolved baseline from params (already narrowed in _build_run_params).
+    baseline = params.get("baseline") or {}
+    metrics["baseline"] = baseline
+
     record: dict[str, Any] = {
         "test_name": test_name,
         "endpoint": endpoint,
         "timestamp": timestamp,
         "server_params": server_cfg.get("server_params"),
         "benchmark_params": params,
+        "baseline": baseline,
         "result": metrics,
         "log_file": str(log_file),
         "Model": model,
@@ -816,70 +826,14 @@ def run_benchmark(
 # ---------------------------------------------------------------------------
 
 
-def _resolve_baseline_value(
-    baseline_raw: Any,
-    *,
-    sweep_index: int | None,
-    max_concurrency: Any = None,
-    request_rate: Any = None,
-) -> Any:
-    """Pick the baseline threshold for this sweep step."""
-    if baseline_raw is None:
-        return 100000
-    if isinstance(baseline_raw, dict):
-        if max_concurrency is not None:
-            for key in (max_concurrency, str(max_concurrency)):
-                if key in baseline_raw:
-                    return baseline_raw[key]
-        if request_rate is not None:
-            for key in (request_rate, str(request_rate)):
-                if key in baseline_raw:
-                    return baseline_raw[key]
-        raise KeyError(
-            f"baseline dict has no key for max_concurrency={max_concurrency!r} "
-            f"or request_rate={request_rate!r}; keys={list(baseline_raw.keys())!r}"
-        )
-    if isinstance(baseline_raw, (list, tuple)):
-        if sweep_index is None:
-            raise ValueError("sweep_index is required when baseline is a list or tuple")
-        return baseline_raw[sweep_index]
-    return baseline_raw
-
-
 def assert_result(
     result: dict[str, Any],
-    params: dict[str, Any],
     num_prompts: int,
-    *,
-    sweep_index: int | None = None,
-    max_concurrency: Any = None,
-    request_rate: Any = None,
-    assert_baseline: bool = True,
 ) -> None:
-    """Assert that benchmark metrics satisfy the configured baselines."""
+    """Assert that the expected number of requests completed."""
     completed = result.get("completed_requests", result.get("completed", 0))
     assert completed == num_prompts, f"Expected {num_prompts} completed requests, got {completed}"
 
-    if not assert_baseline:
-        return
-    if params.get("skip-performance-assertion", False):
-        print("Skipping performance assertions.")
-        return
-
-    for metric, baseline_raw in params.get("baseline", {}).items():
-        current = result.get(metric)
-        assert current is not None, f"Metric '{metric}' not found in result: {list(result.keys())}"
-        threshold = _resolve_baseline_value(
-            baseline_raw,
-            sweep_index=sweep_index,
-            max_concurrency=max_concurrency,
-            request_rate=request_rate,
-        )
-        if "throughput" in metric:
-            assert current >= threshold, f"{metric}: {current:.4f} < baseline {threshold}"
-        else:
-            assert current <= threshold, f"{metric}: {current:.4f} > baseline {threshold}"
-
 
 def _default_benchmark_endpoint_for_task(task: str) -> str:
     """Return the default client-side benchmark endpoint for a diffusion task."""
@@ -921,15 +875,11 @@ def _build_run_params(
     if max_concurrency is not None:
         run_params["max-concurrency"] = max_concurrency
     if "baseline" in params:
-        run_params["baseline"] = {
-            metric: _resolve_baseline_value(
-                baseline_raw,
-                sweep_index=sweep_index,
-                max_concurrency=max_concurrency,
-                request_rate=request_rate,
-            )
-            for metric, baseline_raw in params["baseline"].items()
-        }
+        # Keep all hardware buckets; pick the metric value for this sweep step only.
+        run_params["baseline"] = resolve_baseline_for_sweep(
+            params.get("baseline"),
+            sweep_index=sweep_index,
+        )
     return run_params
 
 
@@ -952,36 +902,30 @@ def _iter_sweep_runs(params: dict[str, Any]) -> list[dict[str, Any]]:
 
     sweep_runs: list[dict[str, Any]] = []
 
-    for i, (request_rate, num_prompts) in enumerate(zip(request_rate_list, num_prompt_list)):
+    for sweep_index, (request_rate, num_prompts) in enumerate(zip(request_rate_list, num_prompt_list)):
         sweep_runs.append(
             {
                 "params": _build_run_params(
                     params,
                     request_rate=request_rate,
                     num_prompts=num_prompts,
-                    sweep_index=i,
+                    sweep_index=sweep_index,
                 ),
                 "num_prompts": num_prompts,
-                "sweep_index": i,
-                "request_rate": request_rate,
-                "max_concurrency": None,
             }
         )
 
-    for i, (max_concurrency, num_prompts) in enumerate(zip(max_concurrency_list, num_prompt_list)):
+    for sweep_index, (max_concurrency, num_prompts) in enumerate(zip(max_concurrency_list, num_prompt_list)):
         sweep_runs.append(
             {
                 "params": _build_run_params(
                     params,
                     max_concurrency=max_concurrency,
                     num_prompts=num_prompts,
-                    sweep_index=i,
                     request_rate="inf",
+                    sweep_index=sweep_index,
                 ),
                 "num_prompts": num_prompts,
-                "sweep_index": i,
-                "request_rate": None,
-                "max_concurrency": max_concurrency,
             }
         )
 
@@ -992,11 +936,9 @@ def _iter_sweep_runs(params: dict[str, Any]) -> list[dict[str, Any]]:
                 "params": _build_run_params(
                     params,
                     num_prompts=default_num_prompts,
+                    sweep_index=0,
                 ),
                 "num_prompts": default_num_prompts,
-                "sweep_index": None,
-                "request_rate": None,
-                "max_concurrency": None,
             }
         )
 
@@ -1012,14 +954,12 @@ def _iter_sweep_runs(params: dict[str, Any]) -> list[dict[str, Any]]:
     paired_benchmark_params,
     indirect=["diffusion_server", "benchmark_params"],
 )
-def test_diffusion_performance_benchmark(diffusion_server, benchmark_params, request):
+def test_diffusion_performance_benchmark(diffusion_server, benchmark_params):
     """Run the diffusion performance benchmark and verify request completion.
 
     One server is started per unique parallel configuration (module scope).
     For each server, all benchmark parameter sets defined in the config JSON
     are executed sequentially; metrics are recorded to the aggregated result file.
-
-    Pass ``--assert-baseline`` to compare ``throughput_qps``, ``latency_mean``, etc. to the JSON ``baseline`` block.
     """
     test_name = benchmark_params["test_name"]
     params = benchmark_params["params"]
@@ -1062,14 +1002,4 @@ def test_diffusion_performance_benchmark(diffusion_server, benchmark_params, req
             print(f"\n  Aggregated results: {_aggregated_result_file_for_source(str(source))}")
         print("=" * 60)
 
-        assert_baseline = request.config.getoption("--assert-baseline", default=False)
-
-        assert_result(
-            result,
-            params,
-            sweep_run["num_prompts"],
-            sweep_index=sweep_run["sweep_index"],
-            max_concurrency=sweep_run["max_concurrency"],
-            request_rate=sweep_run["request_rate"],
-            assert_baseline=assert_baseline,
-        )
+        assert_result(result, sweep_run["num_prompts"])

tests/dfx/perf/tests/test_bagel_vllm_omni.json

@@ -40,10 +40,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.495,
-          "latency_mean": 2.0238,
-          "peak_memory_mb_max": 29393.4286,
-          "peak_memory_mb_mean": 29392.7143
+          "H100": {
+            "throughput_qps": 0.495,
+            "latency_mean": 2.0238,
+            "peak_memory_mb_max": 29393.4286,
+            "peak_memory_mb_mean": 29392.7143
+          }
         }
       }
     ]
@@ -90,10 +92,12 @@
         "enable-negative-prompt": true,
         "num-input-images": 1,
         "baseline": {
-          "throughput_qps": 0.3451,
-          "latency_mean": 2.9031,
-          "peak_memory_mb_max": 30480.2857,
-          "peak_memory_mb_mean": 30479.2857
+          "H100": {
+            "throughput_qps": 0.3451,
+            "latency_mean": 2.9031,
+            "peak_memory_mb_max": 30480.2857,
+            "peak_memory_mb_mean": 30479.2857
+          }
         }
       }
     ]
@@ -139,10 +143,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.5048,
-          "latency_mean": 1.987,
-          "peak_memory_mb_max": 29380.0,
-          "peak_memory_mb_mean": 29380.0
+          "H100": {
+            "throughput_qps": 0.5048,
+            "latency_mean": 1.987,
+            "peak_memory_mb_max": 29380.0,
+            "peak_memory_mb_mean": 29380.0
+          }
         }
       }
     ]
@@ -189,10 +195,12 @@
         "enable-negative-prompt": true,
         "num-input-images": 1,
         "baseline": {
-          "throughput_qps": 0.149,
-          "latency_mean": 6.7157,
-          "peak_memory_mb_max": 30574.0,
-          "peak_memory_mb_mean": 30574.0
+          "H100": {
+            "throughput_qps": 0.149,
+            "latency_mean": 6.7157,
+            "peak_memory_mb_max": 30574.0,
+            "peak_memory_mb_mean": 30574.0
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_boogu_image_edit_vllm_omni.json

@@ -24,10 +24,12 @@
                 "warmup-num-inference-steps": 28,
                 "enable-negative-prompt": true,
                 "baseline": {
-                    "throughput_qps": {"1": 0.0303, "2": 0.0302, "4": 0.0302},
-                    "latency_mean": {"1": 32.6, "2": 62.2, "4": 111.7},
-                    "peak_memory_mb_max": 40500,
-                    "peak_memory_mb_mean": 40500
+                    "H100": {
+                        "throughput_qps": [0.0303, 0.0302, 0.0302],
+                        "latency_mean": [32.6, 62.2, 111.7],
+                        "peak_memory_mb_max": 40500,
+                        "peak_memory_mb_mean": 40500
+                    }
                 }
             }
         ]

tests/dfx/perf/tests/test_boogu_image_vllm_omni.json

@@ -24,10 +24,12 @@
                 "warmup-num-inference-steps": 28,
                 "enable-negative-prompt": true,
                 "baseline": {
-                    "throughput_qps": {"1": 0.0549, "2": 0.0547, "4": 0.0549},
-                    "latency_mean": {"1": 18.0, "2": 34.3, "4": 61.4},
-                    "peak_memory_mb_max": 40500,
-                    "peak_memory_mb_mean": 40500
+                    "H100": {
+                        "throughput_qps": [0.0549, 0.0547, 0.0549],
+                        "latency_mean": [18.0, 34.3, 61.4],
+                        "peak_memory_mb_max": 40500,
+                        "peak_memory_mb_mean": 40500
+                    }
                 }
             }
         ]

tests/dfx/perf/tests/test_cosmos3_vllm_omni.json

@@ -48,9 +48,11 @@
           "guardrails": false
         },
         "baseline": {
-          "throughput_qps": 1.107,
-          "latency_mean": 0.8929,
-          "peak_memory_mb_mean": 80000.0
+          "H100": {
+            "throughput_qps": 1.107,
+            "latency_mean": 0.8929,
+            "peak_memory_mb_mean": 80000.0
+          }
         }
       }
     ]
@@ -115,9 +117,11 @@
           "use_duration_template": false
         },
         "baseline": {
-          "throughput_qps": 0.027,
-          "latency_mean": 36.6824,
-          "peak_memory_mb_mean": 30155.4
+          "H100": {
+            "throughput_qps": 0.027,
+            "latency_mean": 36.6824,
+            "peak_memory_mb_mean": 30155.4
+          }
         }
       }
     ]
@@ -181,9 +185,11 @@
           "guardrails": false
         },
         "baseline": {
-          "throughput_qps": 0.027,
-          "latency_mean": 36.968,
-          "peak_memory_mb_mean": 30443.6
+          "H100": {
+            "throughput_qps": 0.027,
+            "latency_mean": 36.968,
+            "peak_memory_mb_mean": 30443.6
+          }
         }
       }
     ]
@@ -251,9 +257,11 @@
           "condition_video_keep": "first"
         },
         "baseline": {
-          "throughput_qps": 0.027,
-          "latency_mean": 37.2117,
-          "peak_memory_mb_mean": 30135.6
+          "H100": {
+            "throughput_qps": 0.027,
+            "latency_mean": 37.2117,
+            "peak_memory_mb_mean": 30135.6
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_higgs_audio_v3.json

@@ -46,12 +46,14 @@
         "seed_tts_locale": "en",
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            223.3584
-          ],
-          "median_audio_rtf": [
-            0.4393
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              223.3584
+            ],
+            "median_audio_rtf": [
+              0.4393
+            ]
+          }
         }
       },
       {
@@ -70,15 +72,17 @@
         "seed_tts_locale": "en",
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            276.575
-          ],
-          "median_audio_rtf": [
-            0.5073
-          ],
-          "audio_throughput": [
-            27.3728
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              276.575
+            ],
+            "median_audio_rtf": [
+              0.5073
+            ],
+            "audio_throughput": [
+              27.3728
+            ]
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_hunyuan_image3_it2i.json

@@ -140,11 +140,13 @@
         "num-prompts": 96,
         "skip-performance-assertion": true,
         "baseline": {
-          "throughput_qps": 0.1874,
-          "latency_p99": 101.852,
-          "latency_mean": 79.0385,
-          "peak_memory_mb_max": 70000,
-          "peak_memory_mb_mean": 70000
+          "H200": {
+            "throughput_qps": 0.1874,
+            "latency_p99": 101.852,
+            "latency_mean": 79.0385,
+            "peak_memory_mb_max": 70000,
+            "peak_memory_mb_mean": 70000
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json

@@ -36,10 +36,12 @@
         "max-concurrency": 1,
         "skip-performance-assertion": true,
         "baseline": {
-          "throughput_qps": 0.21,
-          "latency_p99": 4.7469,
-          "peak_memory_mb_max": 101100,
-          "peak_memory_mb_mean": 101100
+          "H200": {
+            "throughput_qps": 0.21,
+            "latency_p99": 4.7469,
+            "peak_memory_mb_max": 101100,
+            "peak_memory_mb_mean": 101100
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json

@@ -36,10 +36,12 @@
         "max-concurrency": 1,
         "skip-performance-assertion": true,
         "baseline": {
-          "throughput_qps": 0.20,
-          "latency_p99": 5.1025,
-          "peak_memory_mb_max": 97402,
-          "peak_memory_mb_mean": 97402
+          "H200": {
+            "throughput_qps": 0.2,
+            "latency_p99": 5.1025,
+            "peak_memory_mb_max": 97402,
+            "peak_memory_mb_mean": 97402
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_hunyuan_image_tp4.json

@@ -35,11 +35,13 @@
         "max-concurrency": 1,
         "skip-performance-assertion": true,
         "baseline": {
-          "throughput_qps": 0.213,
-          "latency_mean": 4.6953,
-          "latency_p99": 4.8601,
-          "peak_memory_mb_max": 56912.0,
-          "peak_memory_mb_mean": 56912.0
+          "H200": {
+            "throughput_qps": 0.213,
+            "latency_mean": 4.6953,
+            "latency_p99": 4.8601,
+            "peak_memory_mb_max": 56912.0,
+            "peak_memory_mb_mean": 56912.0
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_lingbot_video_vllm_omni.json

@@ -26,8 +26,7 @@
                 "warmup-requests": 1,
                 "warmup-concurrency": 1,
                 "warmup-num-inference-steps": 2,
-                "enable-negative-prompt": true,
-                "skip-performance-assertion": true
+                "enable-negative-prompt": true
             }
         ]
     }

tests/dfx/perf/tests/test_minicpmo_4_5.json

@@ -0,0 +1,144 @@
+[
+  {
+    "test_name": "test_minicpmo_4_5_challenge",
+    "mark": [
+      {
+        "hardware_marks": {
+          "res": {
+            "cuda": "H100",
+            "npu": "A3"
+          },
+          "num_cards": 1
+        }
+      },
+      "full_model",
+      "omni"
+    ],
+    "server_params": {
+      "model": "openbmb/MiniCPM-o-4_5",
+      "extra_cli_args": [
+        "--deploy-config",
+        "vllm_omni/deploy/minicpmo_4_5.yaml",
+        "--trust-remote-code"
+      ]
+    },
+    "benchmark_params": [
+      {
+        "dataset_name": "seed-tts",
+        "dataset_path": "zhaochenyang20/seed-tts-eval",
+        "backend": "openai-chat-omni",
+        "endpoint": "/v1/chat/completions",
+        "num_prompts": [
+          32,
+          64,
+          128
+        ],
+        "max_concurrency": [
+          1,
+          4,
+          8
+        ],
+        "no_oversample": true,
+        "disable_shuffle": true,
+        "trust_remote_code": true,
+        "seed_tts_locale": "zh",
+        "extra_body": {
+          "modalities": [
+            "text",
+            "audio"
+          ],
+          "chat_template_kwargs": {
+            "enable_thinking": false,
+            "use_tts_template": true
+          }
+        },
+        "percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
+        "baseline": {
+          "A3": {
+            "request_throughput": [
+              0.5383,
+              0.6042,
+              0.7547
+            ],
+            "mean_ttft_ms": [
+              333.2633,
+              533.8656,
+              547.3388
+            ],
+            "mean_e2el_ms": [
+              1857.2154,
+              6600.5101,
+              10450.6811
+            ],
+            "mean_audio_ttfp_ms": [
+              986.4666,
+              3411.0754,
+              3352.7515
+            ],
+            "mean_audio_rtf": [
+              0.4423,
+              1.5734,
+              2.3024
+            ]
+          }
+        }
+      }
+    ]
+  },
+  {
+    "test_name": "test_minicpmo_4_5_a2_challenge",
+    "mark": [
+      {
+        "hardware_marks": {
+          "res": {
+            "npu": "A2"
+          },
+          "num_cards": 1
+        }
+      },
+      "full_model",
+      "omni"
+    ],
+    "server_params": {
+      "model": "openbmb/MiniCPM-o-4_5",
+      "extra_cli_args": [
+        "--deploy-config",
+        "vllm_omni/deploy/minicpmo_4_5.yaml",
+        "--trust-remote-code"
+      ]
+    },
+    "benchmark_params": [
+      {
+        "dataset_name": "seed-tts",
+        "dataset_path": "zhaochenyang20/seed-tts-eval",
+        "backend": "openai-chat-omni",
+        "endpoint": "/v1/chat/completions",
+        "num_prompts": [
+          32,
+          64,
+          128
+        ],
+        "max_concurrency": [
+          1,
+          4,
+          8
+        ],
+        "no_oversample": true,
+        "disable_shuffle": true,
+        "trust_remote_code": true,
+        "seed_tts_locale": "zh",
+        "extra_body": {
+          "modalities": [
+            "text",
+            "audio"
+          ],
+          "chat_template_kwargs": {
+            "enable_thinking": false,
+            "use_tts_template": true
+          }
+        },
+        "percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration"
+      }
+    ]
+  }
+]

tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json

@@ -0,0 +1,69 @@
+[
+  {
+    "test_name": "test_minicpmo_4_5_duplex_seed_tts_challenge",
+    "mark": [
+      {
+        "hardware_marks": {
+          "res": {
+            "cuda": "H100",
+            "npu": "A3"
+          },
+          "num_cards": 1
+        }
+      },
+      "full_model",
+      "omni"
+    ],
+    "server_params": {
+      "model": "openbmb/MiniCPM-o-4_5",
+      "extra_cli_args": [
+        "--deploy-config",
+        "vllm_omni/deploy/minicpmo_4_5_duplex.yaml",
+        "--trust-remote-code"
+      ]
+    },
+    "benchmark_params": [
+      {
+        "dataset_name": "seed-tts",
+        "dataset_path": "zhaochenyang20/seed-tts-eval",
+        "backend": "openai-realtime-tts",
+        "endpoint": "/v1/realtime",
+        "num_prompts": [
+          32
+        ],
+        "max_concurrency": [
+          1
+        ],
+        "no_oversample": true,
+        "disable_shuffle": true,
+        "trust_remote_code": true,
+        "seed_tts_locale": "zh",
+        "seed_tts_turns_per_session": 4,
+        "expected_duplex_audio_turns_per_session": 4,
+        "extra_body": {
+          "save_duplex_request_metrics": true
+        },
+        "percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
+        "baseline": {
+          "A3": {
+            "request_throughput": [
+              0.1594
+            ],
+            "mean_ttft_ms": [
+              274.164
+            ],
+            "mean_e2el_ms": [
+              6250.7696
+            ],
+            "mean_audio_ttfp_ms": [
+              994.422
+            ],
+            "mean_audio_rtf": [
+              0.46163
+            ]
+          }
+        }
+      }
+    ]
+  }
+]

tests/dfx/perf/tests/test_minicpmo_4_5_ready.json

@@ -0,0 +1,81 @@
+[
+  {
+    "test_name": "test_minicpmo_4_5_ready",
+    "mark": [
+      {
+        "hardware_marks": {
+          "res": {
+            "cuda": "H100",
+            "npu": "A3"
+          },
+          "num_cards": 1
+        }
+      },
+      "core_model",
+      "omni"
+    ],
+    "server_params": {
+      "model": "openbmb/MiniCPM-o-4_5",
+      "extra_cli_args": [
+        "--deploy-config",
+        "vllm_omni/deploy/minicpmo_4_5.yaml",
+        "--trust-remote-code"
+      ]
+    },
+    "benchmark_params": [
+      {
+        "dataset_name": "seed-tts",
+        "dataset_path": "zhaochenyang20/seed-tts-eval",
+        "backend": "openai-chat-omni",
+        "endpoint": "/v1/chat/completions",
+        "num_prompts": [
+          10,
+          40
+        ],
+        "max_concurrency": [
+          1,
+          4
+        ],
+        "no_oversample": true,
+        "disable_shuffle": true,
+        "trust_remote_code": true,
+        "seed_tts_locale": "zh",
+        "extra_body": {
+          "modalities": [
+            "text",
+            "audio"
+          ],
+          "chat_template_kwargs": {
+            "enable_thinking": false,
+            "use_tts_template": true
+          }
+        },
+        "percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
+        "baseline": {
+          "H100": {
+            "request_throughput": [
+              0.7803,
+              1.1008
+            ],
+            "mean_ttft_ms": [
+              166.9764,
+              307.7133
+            ],
+            "mean_e2el_ms": [
+              1361.1682,
+              3800.6332
+            ],
+            "mean_audio_ttfp_ms": [
+              635.8099,
+              1615.6591
+            ],
+            "mean_audio_rtf": [
+              0.3151,
+              0.8568
+            ]
+          }
+        }
+      }
+    ]
+  }
+]

tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json

@@ -5,7 +5,8 @@
       {
         "hardware_marks": {
           "res": {
-            "cuda": "H100"
+            "cuda": "H100",
+            "npu": "A3"
           },
           "num_cards": 2
         }
@@ -43,27 +44,29 @@
         "ignore_eos": true,
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            96.4061,
-            140.8147,
-            271.9025,
-            362.2916,
-            507.776
-          ],
-          "mean_e2el_ms": [
-            18507.4582,
-            28365.4956,
-            31907.9613,
-            48161.5829,
-            72630.8601
-          ],
-          "mean_audio_rtf": [
-            0.1175,
-            0.18,
-            0.2359,
-            0.3278,
-            0.6296
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              96.4061,
+              140.8147,
+              271.9025,
+              362.2916,
+              507.776
+            ],
+            "mean_e2el_ms": [
+              18507.4582,
+              28365.4956,
+              31907.9613,
+              48161.5829,
+              72630.8601
+            ],
+            "mean_audio_rtf": [
+              0.1175,
+              0.18,
+              0.2359,
+              0.3278,
+              0.6296
+            ]
+          }
         }
       },
       {
@@ -90,15 +93,17 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            354.0951
-          ],
-          "mean_e2el_ms": [
-            3952.5615
-          ],
-          "mean_audio_rtf": [
-            0.117
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              354.0951
+            ],
+            "mean_e2el_ms": [
+              3952.5615
+            ],
+            "mean_audio_rtf": [
+              0.117
+            ]
+          }
         }
       },
       {
@@ -127,15 +132,17 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            259.6317
-          ],
-          "mean_e2el_ms": [
-            4253.3009
-          ],
-          "mean_audio_rtf": [
-            0.1448
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              259.6317
+            ],
+            "mean_e2el_ms": [
+              4253.3009
+            ],
+            "mean_audio_rtf": [
+              0.1448
+            ]
+          }
         }
       },
       {
@@ -166,15 +173,17 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            457.4453
-          ],
-          "mean_e2el_ms": [
-            5912.5623
-          ],
-          "mean_audio_rtf": [
-            0.1967
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              457.4453
+            ],
+            "mean_e2el_ms": [
+              5912.5623
+            ],
+            "mean_audio_rtf": [
+              0.1967
+            ]
+          }
         }
       },
       {
@@ -205,20 +214,22 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            90.8033,
-            202.1779,
-            326.5864,
-            544.8914,
-            961.3929
-          ],
-          "mean_e2el_ms": [
-            4517.3411,
-            7857.9711,
-            9861.1687,
-            13283.3363,
-            19541.0736
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              90.8033,
+              202.1779,
+              326.5864,
+              544.8914,
+              961.3929
+            ],
+            "mean_e2el_ms": [
+              4517.3411,
+              7857.9711,
+              9861.1687,
+              13283.3363,
+              19541.0736
+            ]
+          }
         }
       },
       {
@@ -250,12 +261,14 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            318.048
-          ],
-          "mean_e2el_ms": [
-            807.6114
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              318.048
+            ],
+            "mean_e2el_ms": [
+              807.6114
+            ]
+          }
         }
       },
       {
@@ -289,12 +302,14 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            186.168
-          ],
-          "mean_e2el_ms": [
-            726.1137
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              186.168
+            ],
+            "mean_e2el_ms": [
+              726.1137
+            ]
+          }
         }
       },
       {
@@ -330,12 +345,14 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            375.5282
-          ],
-          "mean_e2el_ms": [
-            1006.1663
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              375.5282
+            ],
+            "mean_e2el_ms": [
+              1006.1663
+            ]
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen3_omni_multi_replicas.json

@@ -35,10 +35,12 @@
         "ignore_eos": true,
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-            "mean_ttft_ms": [130.5751, 278.6469, 324.7249, 911.6085],
-            "mean_tpot_ms": [7.9854, 11.1209, 16.1507, 20.7485],
-            "mean_audio_ttfp_ms": [486.486, 954.1791, 1289.6514, 2196.8275],
-            "mean_audio_rtf": [0.1896, 0.3962, 0.4098, 0.7322]
+            "H100": {
+                "mean_ttft_ms": [130.5751, 278.6469, 324.7249, 911.6085],
+                "mean_tpot_ms": [7.9854, 11.1209, 16.1507, 20.7485],
+                "mean_audio_ttfp_ms": [486.486, 954.1791, 1289.6514, 2196.8275],
+                "mean_audio_rtf": [0.1896, 0.3962, 0.4098, 0.7322]
+            }
         }
       },
       {
@@ -61,9 +63,11 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-            "mean_ttft_ms": [340.5122],
-            "mean_audio_ttfp_ms": [463.0752],
-            "mean_audio_rtf": [0.1153]
+            "H100": {
+                "mean_ttft_ms": [340.5122],
+                "mean_audio_ttfp_ms": [463.0752],
+                "mean_audio_rtf": [0.1153]
+            }
         }
       },
       {
@@ -88,9 +92,11 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-            "mean_ttft_ms": [236.926],
-            "mean_audio_ttfp_ms": [395.2182],
-            "mean_audio_rtf": [0.1259]
+            "H100": {
+                "mean_ttft_ms": [236.926],
+                "mean_audio_ttfp_ms": [395.2182],
+                "mean_audio_rtf": [0.1259]
+            }
         }
       },
       {
@@ -117,9 +123,11 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-            "mean_ttft_ms": [470.6126],
-            "mean_audio_ttfp_ms": [665.5474],
-            "mean_audio_rtf": [0.166]
+            "H100": {
+                "mean_ttft_ms": [470.6126],
+                "mean_audio_ttfp_ms": [665.5474],
+                "mean_audio_rtf": [0.166]
+            }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json

@@ -5,7 +5,8 @@
       {
         "hardware_marks": {
           "res": {
-            "cuda": "H100"
+            "cuda": "H100",
+            "npu": "A3"
           },
           "num_cards": 2
         }
@@ -43,27 +44,29 @@
         "ignore_eos": true,
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            115.8348,
-            173.1823,
-            446.672,
-            464.5985,
-            3420.0634
-          ],
-          "mean_e2el_ms": [
-            25070.9523,
-            31992.0646,
-            37619.7669,
-            50393.8257,
-            136619.2743
-          ],
-          "mean_audio_rtf": [
-            0.169,
-            0.2111,
-            0.2475,
-            0.3551,
-            0.9266
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              115.8348,
+              173.1823,
+              446.672,
+              464.5985,
+              3420.0634
+            ],
+            "mean_e2el_ms": [
+              25070.9523,
+              31992.0646,
+              37619.7669,
+              50393.8257,
+              136619.2743
+            ],
+            "mean_audio_rtf": [
+              0.169,
+              0.2111,
+              0.2475,
+              0.3551,
+              0.9266
+            ]
+          }
         }
       },
       {
@@ -90,15 +93,17 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            343.1377
-          ],
-          "mean_e2el_ms": [
-            4811.1123
-          ],
-          "mean_audio_rtf": [
-            0.1463
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              343.1377
+            ],
+            "mean_e2el_ms": [
+              4811.1123
+            ],
+            "mean_audio_rtf": [
+              0.1463
+            ]
+          }
         }
       },
       {
@@ -127,15 +132,17 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            245.5446
-          ],
-          "mean_e2el_ms": [
-            5019.1256
-          ],
-          "mean_audio_rtf": [
-            0.1713
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              245.5446
+            ],
+            "mean_e2el_ms": [
+              5019.1256
+            ],
+            "mean_audio_rtf": [
+              0.1713
+            ]
+          }
         }
       },
       {
@@ -166,15 +173,17 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_ttft_ms": [
-            506.1407
-          ],
-          "mean_e2el_ms": [
-            6814.9656
-          ],
-          "mean_audio_rtf": [
-            0.2213
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              506.1407
+            ],
+            "mean_e2el_ms": [
+              6814.9656
+            ],
+            "mean_audio_rtf": [
+              0.2213
+            ]
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json

@@ -51,20 +51,22 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            69.8081,
-            173.0651,
-            271.8108,
-            413.4079,
-            497.2546
-          ],
-          "mean_e2el_ms": [
-            4720.8225,
-            7937.2532,
-            9962.3566,
-            13724.5479,
-            18324.091
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              69.8081,
+              173.0651,
+              271.8108,
+              413.4079,
+              497.2546
+            ],
+            "mean_e2el_ms": [
+              4720.8225,
+              7937.2532,
+              9962.3566,
+              13724.5479,
+              18324.091
+            ]
+          }
         }
       },
       {
@@ -97,12 +99,14 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            25.9477
-          ],
-          "mean_e2el_ms": [
-            517.3155
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              25.9477
+            ],
+            "mean_e2el_ms": [
+              517.3155
+            ]
+          }
         }
       },
       {
@@ -137,12 +141,14 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            26.7056
-          ],
-          "mean_e2el_ms": [
-            551.308
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              26.7056
+            ],
+            "mean_e2el_ms": [
+              551.308
+            ]
+          }
         }
       },
       {
@@ -179,12 +185,14 @@
         },
         "percentile-metrics": "ttft,tpot,itl,e2el",
         "baseline": {
-          "mean_ttft_ms": [
-            27.6906
-          ],
-          "mean_e2el_ms": [
-            596.5284
-          ]
+          "H100": {
+            "mean_ttft_ms": [
+              27.6906
+            ],
+            "mean_e2el_ms": [
+              596.5284
+            ]
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen_image_edit_2511_vllm_omni.json

@@ -34,10 +34,12 @@
         "num-input-images": 2,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0721,
-          "latency_mean": 13.8398,
-          "peak_memory_mb_max": 57610.0,
-          "peak_memory_mb_mean": 57610.0
+          "H100": {
+            "throughput_qps": 0.0721,
+            "latency_mean": 13.8398,
+            "peak_memory_mb_max": 57610.0,
+            "peak_memory_mb_mean": 57610.0
+          }
         }
       },
       {
@@ -52,10 +54,12 @@
         "num-input-images": 2,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0186,
-          "latency_mean": 53.6942,
-          "peak_memory_mb_max": 67790.0,
-          "peak_memory_mb_mean": 67790.0
+          "H100": {
+            "throughput_qps": 0.0186,
+            "latency_mean": 53.6942,
+            "peak_memory_mb_max": 67790.0,
+            "peak_memory_mb_mean": 67790.0
+          }
         }
       }
     ]
@@ -66,7 +70,8 @@
       {
         "hardware_marks": {
           "res": {
-            "cuda": "H100"
+            "cuda": "H100",
+            "npu": "A3"
           },
           "num_cards": 4
         }
@@ -99,10 +104,12 @@
         "num-input-images": 2,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0538,
-          "latency_mean": 18.5591,
-          "peak_memory_mb_max": 56449.4286,
-          "peak_memory_mb_mean": 56449.4286
+          "H100": {
+            "throughput_qps": 0.0538,
+            "latency_mean": 18.5591,
+            "peak_memory_mb_max": 56449.4286,
+            "peak_memory_mb_mean": 56449.4286
+          }
         }
       }
     ]
@@ -156,10 +163,12 @@
         "num-input-images": 2,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.2776,
-          "latency_mean": 3.5938,
-          "peak_memory_mb_max": 58113.7143,
-          "peak_memory_mb_mean": 58075.1429
+          "H100": {
+            "throughput_qps": 0.2776,
+            "latency_mean": 3.5938,
+            "peak_memory_mb_max": 58113.7143,
+            "peak_memory_mb_mean": 58075.1429
+          }
         }
       },
       {
@@ -174,10 +183,12 @@
         "num-input-images": 2,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.1051,
-          "latency_mean": 9.4918,
-          "peak_memory_mb_max": 68107.4286,
-          "peak_memory_mb_mean": 68107.4286
+          "H100": {
+            "throughput_qps": 0.1051,
+            "latency_mean": 9.4918,
+            "peak_memory_mb_max": 68107.4286,
+            "peak_memory_mb_mean": 68107.4286
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen_image_edit_vllm_omni.json

@@ -33,10 +33,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.1129,
-          "latency_mean": 8.8352,
-          "peak_memory_mb_max": 58572.0,
-          "peak_memory_mb_mean": 58572.0
+          "H100": {
+            "throughput_qps": 0.1129,
+            "latency_mean": 8.8352,
+            "peak_memory_mb_max": 58572.0,
+            "peak_memory_mb_mean": 58572.0
+          }
         }
       },
       {
@@ -50,10 +52,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.024,
-          "latency_mean": 41.5686,
-          "peak_memory_mb_max": 66368.0,
-          "peak_memory_mb_mean": 66368.0
+          "H100": {
+            "throughput_qps": 0.024,
+            "latency_mean": 41.5686,
+            "peak_memory_mb_max": 66368.0,
+            "peak_memory_mb_mean": 66368.0
+          }
         }
       }
     ]
@@ -96,10 +100,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0672,
-          "latency_mean": 14.8517,
-          "peak_memory_mb_max": 57076.0,
-          "peak_memory_mb_mean": 57076.0
+          "H100": {
+            "throughput_qps": 0.0672,
+            "latency_mean": 14.8517,
+            "peak_memory_mb_max": 57076.0,
+            "peak_memory_mb_mean": 57076.0
+          }
         }
       }
     ]
@@ -152,10 +158,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.2879,
-          "latency_mean": 3.4701,
-          "peak_memory_mb_max": 58758.0,
-          "peak_memory_mb_mean": 58758.0
+          "H100": {
+            "throughput_qps": 0.2879,
+            "latency_mean": 3.4701,
+            "peak_memory_mb_max": 58758.0,
+            "peak_memory_mb_mean": 58758.0
+          }
         }
       },
       {
@@ -169,10 +177,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.1113,
-          "latency_mean": 8.9656,
-          "peak_memory_mb_max": 67424.0,
-          "peak_memory_mb_mean": 67424.0
+          "H100": {
+            "throughput_qps": 0.1113,
+            "latency_mean": 8.9656,
+            "peak_memory_mb_max": 67424.0,
+            "peak_memory_mb_mean": 67424.0
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen_image_layered_vllm_omni.json

@@ -33,10 +33,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0703,
-          "latency_mean": 14.208,
-          "peak_memory_mb_max": 63225.1429,
-          "peak_memory_mb_mean": 63225.1429
+          "H100": {
+            "throughput_qps": 0.0703,
+            "latency_mean": 14.208,
+            "peak_memory_mb_max": 63225.1429,
+            "peak_memory_mb_mean": 63225.1429
+          }
         }
       },
       {
@@ -50,10 +52,12 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0417,
-          "latency_mean": 23.914,
-          "peak_memory_mb_max": 63225.1429,
-          "peak_memory_mb_mean": 63225.1429
+          "H100": {
+            "throughput_qps": 0.0417,
+            "latency_mean": 23.914,
+            "peak_memory_mb_max": 63225.1429,
+            "peak_memory_mb_mean": 63225.1429
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_qwen_image_vllm_omni.json

@@ -33,9 +33,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.4153,
-          "latency_mean": 2.4065,
-          "peak_memory_mb_mean": 55121.4286
+          "H100": {
+            "throughput_qps": 0.4153,
+            "latency_mean": 2.4065,
+            "peak_memory_mb_mean": 55121.4286
+          }
         }
       },
       {
@@ -49,9 +51,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0423,
-          "latency_mean": 23.5991,
-          "peak_memory_mb_mean": 65118.0
+          "H100": {
+            "throughput_qps": 0.0423,
+            "latency_mean": 23.5991,
+            "peak_memory_mb_mean": 65118.0
+          }
         }
       }
     ]
@@ -91,9 +95,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.4175,
-          "latency_mean": 2.3953,
-          "peak_memory_mb_mean": 55088.5714
+          "H100": {
+            "throughput_qps": 0.4175,
+            "latency_mean": 2.3953,
+            "peak_memory_mb_mean": 55088.5714
+          }
         }
       },
       {
@@ -107,9 +113,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.0423,
-          "latency_mean": 23.5768,
-          "peak_memory_mb_mean": 66012.2857
+          "H100": {
+            "throughput_qps": 0.0423,
+            "latency_mean": 23.5768,
+            "peak_memory_mb_mean": 66012.2857
+          }
         }
       }
     ]
@@ -152,9 +160,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.1214,
-          "latency_mean": 8.2211,
-          "peak_memory_mb_mean": 54501.1429
+          "H100": {
+            "throughput_qps": 0.1214,
+            "latency_mean": 8.2211,
+            "peak_memory_mb_mean": 54501.1429
+          }
         }
       }
     ]
@@ -207,9 +217,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.559,
-          "latency_mean": 1.794,
-          "peak_memory_mb_mean": 55290.5714
+          "H100": {
+            "throughput_qps": 0.559,
+            "latency_mean": 1.794,
+            "peak_memory_mb_mean": 55290.5714
+          }
         }
       },
       {
@@ -223,9 +235,11 @@
         "max-concurrency": 1,
         "enable-negative-prompt": true,
         "baseline": {
-          "throughput_qps": 0.1918,
-          "latency_mean": 5.2022,
-          "peak_memory_mb_mean": 64838.8571
+          "H100": {
+            "throughput_qps": 0.1918,
+            "latency_mean": 5.2022,
+            "peak_memory_mb_mean": 64838.8571
+          }
         }
       }
     ]
@@ -269,9 +283,11 @@
         "warmup-num-inference-steps": 20,
         "enable-negative-prompt": true,
         "baseline": {
-            "throughput_qps": [0.4125, 0.6714, 0.6674, 0.8159],
-            "latency_mean": [2.4215, 2.992, 5.9293, 9.656],
-            "peak_memory_mb_mean": [55147.4286, 55148.0, 55150.4643, 55152.3821]
+            "H100": {
+                "throughput_qps": [0.4125, 0.6714, 0.6674, 0.8159],
+                "latency_mean": [2.4215, 2.992, 5.9293, 9.656],
+                "peak_memory_mb_mean": [55147.4286, 55148.0, 55150.4643, 55152.3821]
+            }
         }
       }
     ]

tests/dfx/perf/tests/test_runner_metadata.py

@@ -1,6 +1,8 @@
 """Tests for DFX runner metadata field exclusion."""
 
 import json
+from contextlib import contextmanager
+from types import SimpleNamespace
 
 import pytest
 
@@ -15,7 +17,7 @@ def test_task_excluded_from_cli_args():
         "backend": "openai-audio-speech",
         "endpoint": "/v1/audio/speech",
         "percentile-metrics": "audio_rtf,audio_ttfp",
-        "baseline": {"mean_audio_rtf": [0.5]},
+        "baseline": {"H100": {"mean_audio_rtf": [0.5]}},
     }
     exclude_keys = {
         "request_rate",
@@ -121,35 +123,6 @@ def test_resolve_pytest_marks_rejects_legacy_object_format():
         )
 
 
-def test_extract_mark_resource_label():
-    from tests.dfx.conftest import extract_mark_resource_label
-
-    assert extract_mark_resource_label(None) == "na"
-    assert (
-        extract_mark_resource_label(
-            [
-                {"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 1}},
-                "full_model",
-            ]
-        )
-        == "H100"
-    )
-    assert (
-        extract_mark_resource_label(
-            [
-                {
-                    "hardware_marks": {
-                        "res": {"cuda": "H100", "rocm": "MI325"},
-                        "num_cards": 2,
-                    }
-                },
-                "diffusion",
-            ]
-        )
-        == "H100-MI325"
-    )
-
-
 def test_resource_label_for_filename():
     from tests.dfx.conftest import resource_label_for_filename
 
@@ -167,16 +140,14 @@ def test_hardware_json_value():
     assert hardware_json_value(None) == ""
 
 
-def test_extract_configs_resource_label(monkeypatch):
-    from tests.dfx.conftest import extract_configs_resource_label, get_runtime_resource_label
+def test_get_runtime_resource_label(monkeypatch):
+    from tests.dfx.conftest import get_runtime_resource_label
 
     monkeypatch.setattr(
         "tests.dfx.conftest._read_runtime_device_name",
         lambda *, device_id=0: "NVIDIA H100 80GB HBM3",
     )
-    get_runtime_resource_label(refresh=True)
-    assert extract_configs_resource_label([]) == "H100"
-    get_runtime_resource_label(refresh=True)
+    assert get_runtime_resource_label(refresh=True) == "H100"
     monkeypatch.setattr(
         "tests.dfx.conftest._read_runtime_device_name",
         lambda *, device_id=0: "Ascend910B2",
@@ -254,8 +225,9 @@ def test_benchmark_param_id_suffix_from_task_eval_phase():
     ]
 
 
-def test_create_paired_omni_benchmark_pytest_params(tmp_path):
+def test_paired_omni_benchmark_reuses_server_and_preserves_case_metadata(tmp_path, monkeypatch):
     from tests.dfx.conftest import create_paired_omni_benchmark_pytest_params
+    from tests.dfx.perf.scripts import run_benchmark
 
     configs = [
         {
@@ -285,3 +257,231 @@ def test_create_paired_omni_benchmark_pytest_params(tmp_path):
     assert bench_row == ("test_tts", 0)
     assert any(m.name == "tts" for m in by_id["test_tts-p0"].marks)
     assert not any(m.name == "tts" for m in by_id["test_omni-p0"].marks)
+
+    omni_p0_server = by_id["test_omni-p0"].values[0]
+    omni_p1_server = by_id["test_omni-p1"].values[0]
+    tts_server = by_id["test_tts-p0"].values[0]
+    assert omni_p0_server == omni_p1_server
+    assert omni_p0_server != tts_server
+
+    events = []
+
+    @contextmanager
+    def fake_start(server_param):
+        events.append(("start", server_param))
+        yield object()
+        events.append(("stop", server_param))
+
+    monkeypatch.setattr(run_benchmark, "_start_omni_server", fake_start)
+    active_context = run_benchmark._SingleActiveContext()
+    try:
+        first = run_benchmark.omni_server.__wrapped__(
+            SimpleNamespace(param=omni_p0_server),
+            active_context,
+        )
+        second = run_benchmark.omni_server.__wrapped__(
+            SimpleNamespace(param=omni_p1_server),
+            active_context,
+        )
+        assert first is second
+        assert events == [("start", omni_p0_server)]
+
+        third = run_benchmark.omni_server.__wrapped__(
+            SimpleNamespace(param=tts_server),
+            active_context,
+        )
+        assert third is not first
+        assert events == [
+            ("start", omni_p0_server),
+            ("stop", omni_p0_server),
+            ("start", tts_server),
+        ]
+    finally:
+        active_context.close()
+
+    assert events == [
+        ("start", omni_p0_server),
+        ("stop", omni_p0_server),
+        ("start", tts_server),
+        ("stop", tts_server),
+    ]
+
+
+def test_is_hardware_nested_baseline():
+    from tests.dfx.conftest import (
+        _RUNTIME_DEVICE_ALIASES,
+        is_hardware_nested_baseline,
+    )
+    from tests.helpers.mark import get_hardware_mark_list
+
+    hardware_marks = get_hardware_mark_list()
+    assert hardware_marks
+    assert {"H100", "L4", "A3", "MI325", "B200"} <= hardware_marks
+    assert "cuda" not in hardware_marks
+    assert "full_model" not in hardware_marks
+
+    # Runtime aliases are a full static list (independent of baseline allowlist).
+    assert "H100" in _RUNTIME_DEVICE_ALIASES
+    assert "A100" in _RUNTIME_DEVICE_ALIASES
+
+    assert is_hardware_nested_baseline(
+        {
+            "H100": {"mean_ttft_ms": [1.0, 2.0], "mean_e2el_ms": [10.0, 20.0]},
+            "A3": {"mean_ttft_ms": [0.9, 1.8], "mean_e2el_ms": [9.0, 18.0]},
+        }
+    )
+    # Custom metric names under a known hardware label are allowed.
+    assert is_hardware_nested_baseline({"H100": {"custom_stage_ms": 12.0, "foo_bar": [1.0, 2.0]}})
+    assert is_hardware_nested_baseline({"A3": {"request_throughput": 0.5}})
+    # Flat metric maps / concurrency-keyed maps are not hardware-nested.
+    assert not is_hardware_nested_baseline({"mean_ttft_ms": [1.0, 2.0], "mean_e2el_ms": [10.0, 20.0]})
+    assert not is_hardware_nested_baseline({"mean_ttft_ms": {"1": 1.0, "32": 2.0}})
+    assert not is_hardware_nested_baseline({})
+    # Alias-only labels (not [hardware-resource] markers) are rejected for baselines.
+    assert not is_hardware_nested_baseline({"A100": {"throughput_qps": 1.0}})
+    # Unknown top-level hardware label is rejected.
+    assert not is_hardware_nested_baseline({"UnknownGPU": {"throughput_qps": 1.0}})
+    # Empty per-hardware metric map is rejected.
+    assert not is_hardware_nested_baseline({"H100": {}})
+
+
+def test_resolve_baseline_for_sweep_keeps_all_hardware_for_one_concurrency():
+    from tests.dfx.conftest import resolve_baseline_for_sweep
+
+    baseline = {
+        "H100": {
+            "mean_ttft_ms": [96.4, 140.8, 271.9, 362.3, 507.8],
+            "mean_e2el_ms": [18507.0, 28365.0, 31907.0, 48161.0, 72630.0],
+        },
+        "A3": {
+            "mean_ttft_ms": [90.0, 130.0, 250.0, 340.0, 480.0],
+            "mean_e2el_ms": [17000.0, 26000.0, 30000.0, 45000.0, 70000.0],
+        },
+    }
+    # max_concurrency=[1,4,8,16,32] -> index 4 is concurrency 32
+    got = resolve_baseline_for_sweep(baseline, sweep_index=4)
+    assert got == {
+        "H100": {"mean_ttft_ms": 507.8, "mean_e2el_ms": 72630.0},
+        "A3": {"mean_ttft_ms": 480.0, "mean_e2el_ms": 70000.0},
+    }
+    # First sweep step keeps both hardware buckets too.
+    got0 = resolve_baseline_for_sweep(baseline, sweep_index=0)
+    assert set(got0) == {"H100", "A3"}
+    assert got0["H100"]["mean_ttft_ms"] == 96.4
+    assert got0["A3"]["mean_ttft_ms"] == 90.0
+
+
+def test_resolve_baseline_for_sweep_rejects_flat_baseline():
+    from tests.dfx.conftest import resolve_baseline_for_sweep
+
+    with pytest.raises(ValueError, match=r"hardware-nested.*pyproject\.toml"):
+        resolve_baseline_for_sweep(
+            {"throughput_qps": [0.4, 0.6], "latency_mean": [1.0, 2.0]},
+            sweep_index=1,
+        )
+
+
+def test_resolve_baseline_for_sweep_rejects_unknown_hardware_label():
+    from tests.dfx.conftest import resolve_baseline_for_sweep
+
+    with pytest.raises(ValueError, match=r"Unknown hardware label\(s\): \['A100'\].*pyproject\.toml"):
+        resolve_baseline_for_sweep(
+            {"A100": {"throughput_qps": 1.0}},
+            sweep_index=0,
+        )
+
+
+def test_resolve_baseline_for_sweep_supports_list_and_scalar_under_hardware():
+    from tests.dfx.conftest import resolve_baseline_for_sweep
+
+    # Sweep-aligned lists under each hardware bucket (canonical form).
+    listed = {
+        "H100": {"throughput_qps": [0.4, 0.6, 0.8], "latency_mean": [1.0, 2.0, 3.0]},
+        "A3": {"throughput_qps": [0.5, 0.7, 0.9], "latency_mean": [0.9, 1.8, 2.7]},
+    }
+    assert resolve_baseline_for_sweep(listed, sweep_index=1) == {
+        "H100": {"throughput_qps": 0.6, "latency_mean": 2.0},
+        "A3": {"throughput_qps": 0.7, "latency_mean": 1.8},
+    }
+
+    # Scalars under hardware stay as-is (single-concurrency cases).
+    scalar = {"H100": {"throughput_qps": 0.5}}
+    assert resolve_baseline_for_sweep(scalar, sweep_index=0) == {"H100": {"throughput_qps": 0.5}}
+    # Custom metric names are preserved.
+    custom = {"H100": {"custom_stage_ms": [10.0, 20.0]}}
+    assert resolve_baseline_for_sweep(custom, sweep_index=1) == {"H100": {"custom_stage_ms": 20.0}}
+    assert resolve_baseline_for_sweep(None) == {}
+    assert resolve_baseline_for_sweep({}) == {}
+
+
+def test_resolve_baseline_value_errors():
+    from tests.dfx.conftest import resolve_baseline_value
+
+    with pytest.raises(ValueError, match="sweep_index"):
+        resolve_baseline_value([1.0, 2.0], sweep_index=None)
+    with pytest.raises(IndexError):
+        resolve_baseline_value([1.0], sweep_index=1)
+    with pytest.raises(TypeError, match="not supported"):
+        resolve_baseline_value({"1": 0.4}, sweep_index=0)
+
+
+def test_diffusion_build_run_params_resolves_baseline_per_sweep(tmp_path, monkeypatch):
+    """``_build_run_params`` / ``_iter_sweep_runs`` narrow list baselines per concurrency."""
+    import importlib
+    import sys
+
+    cfg = tmp_path / "mini_diffusion_perf.json"
+    cfg.write_text(
+        json.dumps(
+            [
+                {
+                    "test_name": "test_mini",
+                    "server_type": "vllm-omni",
+                    "mark": [
+                        {"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 1}},
+                        "diffusion",
+                        "full_model",
+                    ],
+                    "server_params": {"model": "m/mini"},
+                    "benchmark_params": [{"name": "p0", "num-prompts": 1, "max-concurrency": 1}],
+                }
+            ]
+        ),
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(sys, "argv", ["pytest", "--test-config-file", str(cfg)])
+    sys.modules.pop("tests.dfx.perf.scripts.run_diffusion_benchmark", None)
+    from tests.dfx.perf.scripts import run_diffusion_benchmark as rdb
+
+    importlib.reload(rdb)
+
+    params = {
+        "name": "c_sweep",
+        "dataset": "random",
+        "task": "t2i",
+        "num-prompts": [8, 16, 32],
+        "max-concurrency": [1, 8, 32],
+        "baseline": {
+            "H100": {"throughput_qps": [0.1, 0.2, 0.3], "latency_mean": [10.0, 20.0, 30.0]},
+            "A3": {"throughput_qps": [0.11, 0.22, 0.33], "latency_mean": [9.0, 19.0, 29.0]},
+        },
+    }
+    run32 = rdb._build_run_params(
+        params,
+        num_prompts=32,
+        max_concurrency=32,
+        request_rate="inf",
+        sweep_index=2,
+    )
+    assert run32["max-concurrency"] == 32
+    assert run32["baseline"] == {
+        "H100": {"throughput_qps": 0.3, "latency_mean": 30.0},
+        "A3": {"throughput_qps": 0.33, "latency_mean": 29.0},
+    }
+
+    sweeps = rdb._iter_sweep_runs(params)
+    assert len(sweeps) == 3
+    assert sweeps[2]["params"]["max-concurrency"] == 32
+    assert sweeps[2]["params"]["baseline"]["H100"]["throughput_qps"] == 0.3
+    assert "A3" in sweeps[2]["params"]["baseline"]
+    assert not isinstance(sweeps[2]["params"]["baseline"]["H100"]["throughput_qps"], list)

tests/dfx/perf/tests/test_tts.json

@@ -39,12 +39,14 @@
         "seed_tts_locale": "en",
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            141.2791
-          ],
-          "median_audio_rtf": [
-            0.1591
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              141.2791
+            ],
+            "median_audio_rtf": [
+              0.1591
+            ]
+          }
         }
       },
       {
@@ -67,21 +69,23 @@
         "seed_tts_locale": "en",
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            226.747,
-            647.1272,
-            4777.2714
-          ],
-          "median_audio_rtf": [
-            0.3313,
-            0.3642,
-            1.2282
-          ],
-          "audio_throughput": [
-            24.2296,
-            40.3552,
-            43.2572
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              226.747,
+              647.1272,
+              4777.2714
+            ],
+            "median_audio_rtf": [
+              0.3313,
+              0.3642,
+              1.2282
+            ],
+            "audio_throughput": [
+              24.2296,
+              40.3552,
+              43.2572
+            ]
+          }
         }
       },
       {
@@ -102,9 +106,11 @@
         "seed_tts_wer_eval": true,
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_audio_rtf": [
-            0.45
-          ]
+          "H100": {
+            "mean_audio_rtf": [
+              0.45
+            ]
+          }
         }
       }
     ]
@@ -153,12 +159,14 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            46.8434
-          ],
-          "median_audio_rtf": [
-            0.1367
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              46.8434
+            ],
+            "median_audio_rtf": [
+              0.1367
+            ]
+          }
         }
       },
       {
@@ -186,21 +194,23 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            75.3341,
-            713.7727,
-            5941.7179
-          ],
-          "median_audio_rtf": [
-            0.1859,
-            0.3148,
-            1.1662
-          ],
-          "audio_throughput": [
-            38.3858,
-            55.3523,
-            68.7032
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              75.3341,
+              713.7727,
+              5941.7179
+            ],
+            "median_audio_rtf": [
+              0.1859,
+              0.3148,
+              1.1662
+            ],
+            "audio_throughput": [
+              38.3858,
+              55.3523,
+              68.7032
+            ]
+          }
         }
       },
       {
@@ -226,9 +236,11 @@
         "seed_tts_wer_eval": true,
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_audio_rtf": [
-            0.35
-          ]
+          "H100": {
+            "mean_audio_rtf": [
+              0.35
+            ]
+          }
         }
       },
       {
@@ -252,15 +264,17 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            56.746
-          ],
-          "median_audio_rtf": [
-            0.1571
-          ],
-          "audio_throughput": [
-            11.1226
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              56.746
+            ],
+            "median_audio_rtf": [
+              0.1571
+            ],
+            "audio_throughput": [
+              11.1226
+            ]
+          }
         }
       },
       {
@@ -283,12 +297,14 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            46.3479
-          ],
-          "median_audio_rtf": [
-            0.1352
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              46.3479
+            ],
+            "median_audio_rtf": [
+              0.1352
+            ]
+          }
         }
       },
       {
@@ -315,21 +331,23 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            75.5967,
-            587.5877,
-            4960.3726
-          ],
-          "median_audio_rtf": [
-            0.1871,
-            0.3143,
-            1.1703
-          ],
-          "audio_throughput": [
-            38.9378,
-            55.1846,
-            71.5708
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              75.5967,
+              587.5877,
+              4960.3726
+            ],
+            "median_audio_rtf": [
+              0.1871,
+              0.3143,
+              1.1703
+            ],
+            "audio_throughput": [
+              38.9378,
+              55.1846,
+              71.5708
+            ]
+          }
         }
       },
       {
@@ -352,15 +370,17 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            55.5029
-          ],
-          "median_audio_rtf": [
-            0.1508
-          ],
-          "audio_throughput": [
-            9.4885
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              55.5029
+            ],
+            "median_audio_rtf": [
+              0.1508
+            ],
+            "audio_throughput": [
+              9.4885
+            ]
+          }
         }
       }
     ]
@@ -402,12 +422,14 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            82.1886
-          ],
-          "median_audio_rtf": [
-            0.1241
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              82.1886
+            ],
+            "median_audio_rtf": [
+              0.1241
+            ]
+          }
         }
       },
       {
@@ -429,15 +451,17 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            160.9168
-          ],
-          "median_audio_rtf": [
-            0.1907
-          ],
-          "audio_throughput": [
-            39.2332
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              160.9168
+            ],
+            "median_audio_rtf": [
+              0.1907
+            ],
+            "audio_throughput": [
+              39.2332
+            ]
+          }
         }
       },
       {
@@ -459,15 +483,17 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            100.1265
-          ],
-          "median_audio_rtf": [
-            0.1437
-          ],
-          "audio_throughput": [
-            11.7605
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              100.1265
+            ],
+            "median_audio_rtf": [
+              0.1437
+            ],
+            "audio_throughput": [
+              11.7605
+            ]
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_voxcpm2.json

@@ -37,12 +37,14 @@
         "seed_tts_locale": "en",
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            137.645
-          ],
-          "median_audio_rtf": [
-            0.0497
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              137.645
+            ],
+            "median_audio_rtf": [
+              0.0497
+            ]
+          }
         }
       },
       {
@@ -61,15 +63,17 @@
         "seed_tts_locale": "en",
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            475.3357
-          ],
-          "median_audio_rtf": [
-            0.1243
-          ],
-          "audio_throughput": [
-            36.6358
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              475.3357
+            ],
+            "median_audio_rtf": [
+              0.1243
+            ],
+            "audio_throughput": [
+              36.6358
+            ]
+          }
         }
       },
       {
@@ -90,9 +94,11 @@
         "seed_tts_wer_eval": true,
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_audio_rtf": [
-            0.45
-          ]
+          "H100": {
+            "mean_audio_rtf": [
+              0.45
+            ]
+          }
         }
       },
       {
@@ -114,12 +120,14 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            79.6927
-          ],
-          "median_audio_rtf": [
-            0.0423
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              79.6927
+            ],
+            "median_audio_rtf": [
+              0.0423
+            ]
+          }
         }
       },
       {
@@ -141,15 +149,17 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            262.6212
-          ],
-          "median_audio_rtf": [
-            0.0753
-          ],
-          "audio_throughput": [
-            103.5843
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              262.6212
+            ],
+            "median_audio_rtf": [
+              0.0753
+            ],
+            "audio_throughput": [
+              103.5843
+            ]
+          }
         }
       },
       {
@@ -171,15 +181,17 @@
         },
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "median_audio_ttfp_ms": [
-            80.1095
-          ],
-          "median_audio_rtf": [
-            0.0479
-          ],
-          "audio_throughput": [
-            17.2918
-          ]
+          "H100": {
+            "median_audio_ttfp_ms": [
+              80.1095
+            ],
+            "median_audio_rtf": [
+              0.0479
+            ],
+            "audio_throughput": [
+              17.2918
+            ]
+          }
         }
       },
       {
@@ -203,9 +215,11 @@
         "seed_tts_wer_eval": true,
         "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
         "baseline": {
-          "mean_audio_rtf": [
-            0.35
-          ]
+          "H100": {
+            "mean_audio_rtf": [
+              0.35
+            ]
+          }
         }
       }
     ]

tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json

@@ -42,9 +42,11 @@
           }
         ],
         "baseline": {
-          "throughput_qps": 0.0434,
-          "latency_mean": 23.0654,
-          "peak_memory_mb_mean": 76360.0
+          "H100": {
+            "throughput_qps": 0.0434,
+            "latency_mean": 23.0654,
+            "peak_memory_mb_mean": 76360.0
+          }
         }
       }
     ]
@@ -55,7 +57,8 @@
       {
         "hardware_marks": {
           "res": {
-            "cuda": "H100"
+            "cuda": "H100",
+            "npu": "A3"
           },
           "num_cards": 2
         }
@@ -96,9 +99,11 @@
           }
         ],
         "baseline": {
-          "throughput_qps": 0.0606,
-          "latency_mean": 16.4928,
-          "peak_memory_mb_mean": 47164.8
+          "H100": {
+            "throughput_qps": 0.0606,
+            "latency_mean": 16.4928,
+            "peak_memory_mb_mean": 47164.8
+          }
         }
       },
       {
@@ -121,9 +126,11 @@
           }
         ],
         "baseline": {
-          "throughput_qps": 0.0124,
-          "latency_mean": 80.7784,
-          "peak_memory_mb_mean": 55374.3429
+          "H100": {
+            "throughput_qps": 0.0124,
+            "latency_mean": 80.7784,
+            "peak_memory_mb_mean": 55374.3429
+          }
         }
       }
     ]

tests/dfx/reliability/invalid_param_test/test_invalid_minicpmo_4_5_omni.py

@@ -14,18 +14,15 @@ from tests.helpers.stage_config import get_deploy_config_path
 
 pytestmark = [pytest.mark.slow, pytest.mark.omni, pytest.mark.full_model]
 
-_MINICPMO_DEPLOY = get_deploy_config_path("minicpmo_4_5_batching.yaml")
+_MINICPMO_DEPLOY = get_deploy_config_path("minicpmo_4_5_2gpu.yaml")
 
 _MINICPMO_SERVER = [
     pytest.param(
         OmniServerParams(
             model="openbmb/MiniCPM-o-4_5",
             stage_config_path=_MINICPMO_DEPLOY,
-            use_stage_cli=True,
-            server_args=[
-                "--trust-remote-code",
-                "--no-async-chunk",
-            ],
+            use_stage_cli=False,
+            server_args=["--trust-remote-code"],
         ),
         id="minicpmo_4_5",
     )
@@ -51,12 +48,12 @@ def _prompt() -> str:
     return "What is the capital of China? Answer in 20 words."
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", _MINICPMO_SERVER, indirect=True)
 def test_invalid_audio_format_rejected(omni_server: OmniServerParams, openai_client: OpenAIClientHandler) -> None:
     """
     Test that invalid audio format is properly rejected.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text + invalid audio format
     Output Modal: error
     """

tests/distributed/omni_connectors/test_chunk_transfer_adapter.py

@@ -71,6 +71,30 @@ def test_streaming_payload_can_replace_placeholder_prompt(mocker: MockerFixture)
     request.update_block_hashes.assert_called_once_with()
 
 
+def test_streaming_payload_can_append_exact_prompt_length(mocker: MockerFixture) -> None:
+    request = SimpleNamespace(
+        _all_token_ids=[0, 0, 7, 8],
+        _output_token_ids=[7, 8],
+        prompt_token_ids=[0, 0],
+        num_computed_tokens=4,
+        num_prompt_tokens=2,
+        update_block_hashes=mocker.Mock(),
+    )
+    payload = {
+        "ids": {"prompt": [1, 2, 3]},
+        "meta": {"next_stage_prompt_len": 3},
+    }
+
+    construct_next_stage_streaming_input_prompt(payload, request)
+
+    assert request.prompt_token_ids == [0, 0, 7, 8, 0, 0, 0]
+    assert request._all_token_ids == [0, 0, 7, 8, 0, 0, 0]
+    assert request._output_token_ids == []
+    assert request.num_computed_tokens == 4
+    assert request.num_prompt_tokens == 7
+    request.update_block_hashes.assert_called_once_with()
+
+
 @pytest.fixture
 def build_adapter(monkeypatch, mocker: MockerFixture):
     def _build(
@@ -252,6 +276,37 @@ def test_save_async_uses_confirmed_tokens_for_async_scheduler_watermark(build_ad
     assert len(adapter._pending_save_reqs) == 1
 
 
+def test_send_single_request_terminal_chunk_still_flushes_processor(build_adapter, monkeypatch):
+    """A terminal stop is not a segment boundary (#5383), but the producer-side
+    processor must still receive the flush signal on the terminal chunk.
+    Passing only ``is_segment_finished`` starved processors of their final
+    accumulated payload once terminal stops stopped setting it (#5413: the
+    downstream stage got ``meta.finished`` with the tail data missing).
+    """
+    adapter, connector = build_adapter(stage_id=0)
+    request = _req("req-terminal", RequestStatus.FINISHED_STOPPED, external_req_id="ext-terminal")
+
+    seen_flush_flags = []
+
+    def recording_processor(**kwargs):
+        seen_flush_flags.append(kwargs["is_finished"])
+        return OmniPayloadStruct(
+            codes=CodesStruct(audio=torch.tensor([1, 2, 3], dtype=torch.long)),
+        )
+
+    adapter.custom_process_next_stage_input_func = recording_processor
+    monkeypatch.setattr(adapter, "cleanup", lambda *a, **kw: None)
+
+    adapter._send_single_request(
+        {"multimodal_output": None, "request": request, "is_finished": True, "is_segment_finished": False}
+    )
+
+    assert seen_flush_flags == [True]
+    sent_payload = connector.put.call_args.kwargs["data"]
+    assert bool(sent_payload.meta.finished.item()) is True
+    assert bool(sent_payload.meta.is_segment_finished.item()) is False
+
+
 def test_send_single_request_struct_without_meta_does_not_crash(build_adapter, monkeypatch):
     """Producer may return a struct with ``meta=None`` (e.g. payload that
     carries only ``embed`` or ``codes``). The sender's ``meta is not None``
@@ -314,6 +369,23 @@ def test_send_single_request_struct_preserves_segment_finished(build_adapter, mo
     assert sent_payload.meta.is_segment_finished.item() is True
 
 
+def test_send_single_request_respects_processor_receiver_boundary(build_adapter, monkeypatch):
+    adapter, connector = build_adapter(stage_id=1)
+    request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
+
+    adapter.custom_process_next_stage_input_func = lambda **kwargs: OmniPayloadStruct(
+        meta=MetaStruct(is_segment_finished=torch.tensor(False, dtype=torch.bool))
+    )
+    monkeypatch.setattr(adapter, "cleanup", lambda *a, **kw: None)
+
+    adapter._send_single_request(
+        {"multimodal_output": None, "request": request, "is_finished": False, "is_segment_finished": True}
+    )
+
+    sent_payload = connector.put.call_args.kwargs["data"]
+    assert sent_payload.meta.is_segment_finished.item() is False
+
+
 def test_save_async_skips_stale_resumable_chunk_until_dedup_is_reset(build_adapter):
     adapter, _ = build_adapter(stage_id=1)
     request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
@@ -463,7 +535,7 @@ def test_fifo_promotion(build_adapter):
     adapter.process_pending_chunks(waiting_queue, running_queue)
 
     assert list(adapter._active_streams) == ["req-1", "req-2"]
-    assert waiting_queue == reqs[2:]
+    assert waiting_queue == []
     assert reqs[0].status == RequestStatus.WAITING_FOR_CHUNK
     assert reqs[1].status == RequestStatus.WAITING_FOR_CHUNK
     assert reqs[2].status == RequestStatus.WAITING
@@ -473,6 +545,7 @@ def test_fifo_promotion(build_adapter):
     # (commit c4d95fd9 — otherwise the terminal chunk deadlocks at c=8 K=2).
     # Simulate it here so promotion can pick up the freed slot.
     adapter._evict_finished_active_streams({"req-1"})
+    adapter.restore_queues(waiting_queue, running_queue)
     adapter.process_pending_chunks(waiting_queue, running_queue)
 
     assert list(adapter._active_streams) == ["req-2", "req-3"]
@@ -481,6 +554,22 @@ def test_fifo_promotion(build_adapter):
     assert reqs[2].status == RequestStatus.WAITING_FOR_CHUNK
 
 
+def test_non_active_waiting_request_is_held_off_scheduler(build_adapter):
+    adapter, _ = build_adapter(stage_id=2, max_num_seqs=2, active_stream_window=1)
+    active = _req("req-active", RequestStatus.WAITING)
+    non_active = _req("req-non-active", RequestStatus.WAITING)
+    waiting_queue = DummyWaitingQueue([active, non_active])
+    running_queue = []
+
+    adapter.process_pending_chunks(waiting_queue, running_queue)
+
+    assert waiting_queue == []
+    assert active.status == RequestStatus.WAITING_FOR_CHUNK
+    assert non_active.status == RequestStatus.WAITING
+    assert list(adapter._pending_load_reqs) == [active]
+    assert list(adapter.waiting_for_chunk_waiting_requests) == [active, non_active]
+
+
 def test_legacy_k0(build_adapter):
     adapter, _ = build_adapter(stage_id=1, max_num_seqs=1, active_stream_window=0)
     waiting_req = _req("waiting", RequestStatus.WAITING)
@@ -509,12 +598,13 @@ def test_finished_releases_slot(build_adapter):
 
     adapter.process_pending_chunks(waiting_queue, running_queue)
     assert list(adapter._active_streams) == ["req-1"]
-    assert waiting_queue == [req_2]
+    assert waiting_queue == []
 
     adapter.finished_requests.add("req-1")
     # Eviction is deferred to postprocess_scheduler_output in the runtime path
     # (commit c4d95fd9). Simulate it so promotion can pick up the freed slot.
     adapter._evict_finished_active_streams({"req-1"})
+    adapter.restore_queues(waiting_queue, running_queue)
     adapter.process_pending_chunks(waiting_queue, running_queue)
 
     assert list(adapter._active_streams) == ["req-2"]

tests/e2e/accuracy/minicpmo_4_5/__init__.py

@@ -0,0 +1 @@
+# SPDX-License-Identifier: Apache-2.0

tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py

@@ -0,0 +1,296 @@
+# SPDX-License-Identifier: Apache-2.0
+"""MiniCPM-o 4.5 Daily-Omni + Seed-TTS + Video-MME accuracy regression coverage.
+
+Daily-Omni settings follow the MiniCPM interleaved AV recipe (``minicpm-interleave``,
+``temperature=0``, text modalities with ``use_tts_template``, and server
+``--interleave-mm-strings`` + 1fps / 128-frame media-io kwargs, also pinned in
+``minicpmo_4_5.yaml``). Baseline **79.5**; gate **77.5** (drop ≤ 2pp).
+
+Video-MME settings follow OpenBMB OmniEvalKit MiniCPM ``videomme`` (w/o subs):
+``minicpm-frames``, ``max_frames=96``, ``temperature=0``, ``output_len=128``.
+Baseline **69.0**; gate **67.0** (drop ≤ 2pp).
+
+Seed-TTS (zh, seed-tts-eval protocol): WER baseline **1.414%** / gate **≤ 1.56%**
+(+10%).
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import json
+import os
+from pathlib import Path
+
+import pytest
+
+from tests.e2e.accuracy.qwen3_omni import run_qwen_omni_acc_benchmark as _acc_bench
+from tests.e2e.accuracy.qwen3_omni.qwen3_omni_acc_bench_core import (
+    build_acc_benchmark_cli_argv,
+    find_vllm_cli,
+)
+from tests.helpers.mark import hardware_test
+from tests.helpers.runtime import OmniServerParams
+from tests.helpers.stage_config import get_deploy_config_path
+
+_MODEL = os.environ.get("VLLM_TEST_MINICPMO_4_5_MODEL", "openbmb/MiniCPM-o-4_5")
+_DEPLOY_CONFIG = get_deploy_config_path("minicpmo_4_5.yaml")
+_RESULT_DIR = Path(
+    os.environ.get(
+        "ACC_BENCH_RESULT_DIR",
+        str(Path(__file__).resolve().parent / "results"),
+    )
+)
+
+# Accuracy gates: baseline minus allowed drop (≤2pp / ≤10% WER).
+_MIN_DAILY_OMNI_ACCURACY = 0.775  # 79.5 - 2pp
+_MIN_VIDEOMME_ACCURACY = float(os.environ.get("ACC_BENCH_MIN_VIDEOMME_ACCURACY", "0.67"))
+# jiwer reports a 0-1 ratio; table 1.414 / 1.56 is percent → 0.0156.
+_MAX_SEED_TTS_MEAN_WER = 0.0156
+# Full Seed-TTS zh split (seed-tts-eval/zh/meta.lst) is 2020 rows.
+_SEED_TTS_LOCALE = "zh"
+_SEED_TTS_NUM_PROMPTS = 2020
+# Match the validated Daily-Omni / Video-MME client body from the MiniCPM run scripts.
+_DAILY_EXTRA_BODY = {
+    "modalities": ["text"],
+    "chat_template_kwargs": {"enable_thinking": False},
+}
+_VIDEOMME_EXTRA_BODY = {
+    "modalities": ["text"],
+    "chat_template_kwargs": {"enable_thinking": False},
+}
+_SEED_EXTRA_BODY = {
+    "modalities": ["text", "audio"],
+    "chat_template_kwargs": {
+        "enable_thinking": False,
+        "use_tts_template": True,
+    },
+}
+# Server flags required for MiniCPM interleaved image/audio packs (Daily-Omni only).
+# Do not reuse these for Seed-TTS: ``--interleave-mm-strings`` + TTS ref_audio can trip
+# msgspec ValidationError and kill the orchestrator mid-suite.
+# Also pinned in ``minicpmo_4_5.yaml`` stage 0 so they survive config-path routing.
+_DAILY_OMNI_SERVER_ARGS = [
+    "--trust-remote-code",
+    "--interleave-mm-strings",
+    "--media-io-kwargs",
+    '{"video":{"fps":1,"num_frames":128}}',
+]
+# Video-MME ``minicpm-frames`` sends sampled frames as image_url (no AV interleave).
+# Allow local ``file://`` frame-cache URLs so CI does not need megabyte-scale base64
+# payloads; the autouse fixture still appends ``--videomme-inline-local-video`` when
+# the allowlist is unavailable.
+_VIDEOMME_SERVER_ARGS = [
+    "--trust-remote-code",
+    "--allowed-local-media-path",
+    os.environ.get("VLLM_TEST_ALLOWED_LOCAL_MEDIA_PATH", "/"),
+]
+_SEED_TTS_SERVER_ARGS = [
+    "--trust-remote-code",
+]
+
+pytestmark = [pytest.mark.full_model, pytest.mark.omni]
+
+# Evaluated at setup time, before the ``omni_server`` fixture boots a server. An
+# ``importorskip`` inside the test body would skip only after paying several minutes
+# of startup, and the resulting skip is easy to miss in a nightly log.
+_missing_daily_omni_deps = [name for name in ("datasets", "huggingface_hub") if importlib.util.find_spec(name) is None]
+requires_daily_omni_deps = pytest.mark.skipif(
+    bool(_missing_daily_omni_deps),
+    reason=f"Daily-Omni accuracy bench needs {', '.join(_missing_daily_omni_deps)}",
+)
+
+daily_test_params = [
+    OmniServerParams(
+        model=_MODEL,
+        stage_config_path=_DEPLOY_CONFIG,
+        server_args=list(_DAILY_OMNI_SERVER_ARGS),
+    )
+]
+
+videomme_test_params = [
+    OmniServerParams(
+        model=_MODEL,
+        stage_config_path=_DEPLOY_CONFIG,
+        server_args=list(_VIDEOMME_SERVER_ARGS),
+    )
+]
+
+seed_test_params = [
+    OmniServerParams(
+        model=_MODEL,
+        stage_config_path=_DEPLOY_CONFIG,
+        server_args=list(_SEED_TTS_SERVER_ARGS),
+    )
+]
+
+
+def _require_vllm_cli() -> None:
+    try:
+        find_vllm_cli()
+    except FileNotFoundError as exc:
+        pytest.skip(str(exc))
+
+
+def _optional_local_videomme_dataset_path() -> str | None:
+    """Return an explicit local Video-MME root, or ``None`` to use the Hub default.
+
+    Same rule as Daily-Omni / Seed-TTS: only an env-provided existing directory counts as
+    local. Hard-coded workspace paths are intentionally not probed so CI defaults to the
+    Hub id in ``VIDEOMME_DEFAULT_HF_REPO`` (``lmms-eval/Video-MME``).
+    """
+    for key in ("VLLM_VIDEOMME_DATASET_PATH", "VIDEOMME_ROOT"):
+        raw = os.environ.get(key, "").strip()
+        if not raw:
+            continue
+        path = Path(raw).expanduser()
+        if path.is_absolute() and path.is_dir():
+            return str(path.resolve())
+    return None
+
+
+@pytest.fixture(autouse=True)
+def _inline_local_media_when_needed(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Inline Hub/local media when the server has no usable local-media allowlist.
+
+    Daily-Omni always inlines (matches the historical Qwen/MiniCPM accuracy fixture).
+    Video-MME prefers ``file://`` via ``--allowed-local-media-path`` (see
+    ``_VIDEOMME_SERVER_ARGS``); force inline only when ``VLLM_VIDEOMME_FORCE_INLINE=1``.
+    """
+    original_daily = _acc_bench.daily_omni_bench_argv
+    original_videomme = _acc_bench.videomme_bench_argv
+
+    def _wrap_daily() -> list[str]:
+        argv = list(original_daily())
+        if "--daily-omni-inline-local-video" not in argv:
+            argv.append("--daily-omni-inline-local-video")
+        return argv
+
+    def _wrap_videomme() -> list[str]:
+        argv = list(original_videomme())
+        force_inline = os.environ.get("VLLM_VIDEOMME_FORCE_INLINE", "").strip().lower() in {
+            "1",
+            "true",
+            "yes",
+        }
+        if force_inline and "--videomme-inline-local-video" not in argv:
+            argv.append("--videomme-inline-local-video")
+        return argv
+
+    monkeypatch.setattr(_acc_bench, "daily_omni_bench_argv", _wrap_daily)
+    monkeypatch.setattr(_acc_bench, "videomme_bench_argv", _wrap_videomme)
+    monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+
+@requires_daily_omni_deps
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", daily_test_params, indirect=True)
+def test_minicpmo_4_5_daily_omni_accuracy_bench(omni_server) -> None:
+    _require_vllm_cli()
+
+    argv = build_acc_benchmark_cli_argv(
+        omni_server,
+        skip_seed=True,
+        skip_daily=False,
+        skip_videomme=True,
+    )
+    argv.extend(
+        [
+            "--result-dir",
+            str(_RESULT_DIR),
+            "--min-daily-omni-accuracy",
+            str(_MIN_DAILY_OMNI_ACCURACY),
+            "--temperature",
+            "0",
+            "--output-len",
+            "512",
+            "--daily-omni-input-mode",
+            "all",
+            "--daily-omni-pack-mode",
+            "minicpm-interleave",
+            "--daily-extra-body-json",
+            json.dumps(_DAILY_EXTRA_BODY, separators=(",", ":")),
+            "--trust-remote-code",
+        ]
+    )
+
+    assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0
+
+
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", videomme_test_params, indirect=True)
+def test_minicpmo_4_5_videomme_accuracy_bench(omni_server) -> None:
+    """Gate MiniCPM-o 4.5 Video-MME (w/o subs) overall accuracy at the OmniEvalKit recipe."""
+    _require_vllm_cli()
+    pytest.importorskip("datasets")
+    pytest.importorskip("huggingface_hub")
+    pytest.importorskip("av")
+
+    dataset_path = _optional_local_videomme_dataset_path()
+    num_prompts = int(os.environ.get("ACC_BENCH_VIDEOMME_NUM_PROMPTS", "2700"))
+    max_concurrency = int(os.environ.get("ACC_BENCH_VIDEOMME_MAX_CONCURRENCY", "4"))
+
+    argv = build_acc_benchmark_cli_argv(
+        omni_server,
+        skip_seed=True,
+        skip_daily=True,
+        skip_videomme=False,
+        num_prompts=num_prompts,
+        max_concurrency=max_concurrency,
+    )
+    argv.extend(
+        [
+            "--result-dir",
+            str(_RESULT_DIR),
+            "--min-videomme-accuracy",
+            str(_MIN_VIDEOMME_ACCURACY),
+            "--temperature",
+            "0",
+            "--output-len",
+            "128",
+            "--videomme-pack-mode",
+            "minicpm-frames",
+            "--videomme-max-frames",
+            "96",
+            "--videomme-duration",
+            os.environ.get("ACC_BENCH_VIDEOMME_DURATION", "all"),
+            "--videomme-extra-body-json",
+            json.dumps(_VIDEOMME_EXTRA_BODY, separators=(",", ":")),
+            "--trust-remote-code",
+        ]
+    )
+    # Absolute local root only; otherwise --videomme-repo (Hub) from build_acc_benchmark_cli_argv.
+    if dataset_path is not None:
+        argv.extend(["--videomme-dataset-path", dataset_path])
+
+    assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0
+
+
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", seed_test_params, indirect=True)
+def test_minicpmo_4_5_seed_tts_wer_bench(omni_server) -> None:
+    _require_vllm_cli()
+    pytest.importorskip("huggingface_hub")
+
+    argv = build_acc_benchmark_cli_argv(
+        omni_server,
+        skip_seed=False,
+        skip_daily=True,
+        skip_videomme=True,
+        num_prompts=_SEED_TTS_NUM_PROMPTS,
+        max_concurrency=4,
+    )
+    argv.extend(
+        [
+            "--result-dir",
+            str(_RESULT_DIR),
+            "--max-seed-tts-mean-wer",
+            str(_MAX_SEED_TTS_MEAN_WER),
+            "--seed-tts-locale",
+            _SEED_TTS_LOCALE,
+            "--seed-extra-body-json",
+            json.dumps(_SEED_EXTRA_BODY, separators=(",", ":")),
+            "--trust-remote-code",
+        ]
+    )
+
+    assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0

tests/e2e/accuracy/qwen3_omni/qwen3_omni_acc_bench_core.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-"""Shared helpers for Qwen3-Omni Daily-Omni / Seed-TTS ``vllm bench serve --omni`` accuracy runs.
+"""Shared helpers for Omni Daily-Omni / Seed-TTS / Video-MME ``vllm bench serve --omni`` accuracy runs.
 
 Local dataset paths are **optional**. When ``VLLM_DAILY_OMNI_QA_JSON`` + ``VLLM_DAILY_OMNI_VIDEO_DIR``
 point to existing files, those are used with inline video. Otherwise the benchmark falls back to
@@ -9,6 +9,10 @@ bench request that needs media downloads ``Videos.tar`` from the Hub when no vid
 Similarly for Seed-TTS: a local directory wins; otherwise ``--dataset-path`` uses the Hub id
 and ``huggingface_hub.snapshot_download`` inside ``resolve_seed_tts_root`` pulls files on demand.
 
+Video-MME follows the same pattern: an existing local directory via ``VLLM_VIDEOMME_DATASET_PATH`` /
+``VIDEOMME_ROOT`` (or ``--videomme-dataset-path`` / ``--dataset-path``) wins; otherwise the Hub id
+``lmms-eval/Video-MME`` is used and ``resolve_videomme_root`` downloads parquet + video archives.
+
 Use :func:`build_acc_benchmark_cli_argv` to assemble ``argv`` for a live Omni server (host/port/model
 and small bench defaults) before ``parse_args`` / ``run_acc_benchmark`` in the accuracy driver.
 """
@@ -19,11 +23,13 @@ import json
 import os
 import shutil
 import subprocess
+import sys
 from pathlib import Path
 from typing import Any, Protocol
 
 DEFAULT_DAILY_OMNI_HF_REPO = "liarliar/Daily-Omni"
 DEFAULT_SEED_TTS_HF_REPO = "zhaochenyang20/seed-tts-eval"
+DEFAULT_VIDEOMME_HF_REPO = "lmms-eval/Video-MME"
 
 
 class OmniBenchServerEndpoint(Protocol):
@@ -39,6 +45,7 @@ def build_acc_benchmark_cli_argv(
     *,
     skip_seed: bool,
     skip_daily: bool,
+    skip_videomme: bool = True,
     num_prompts: int | None = None,
     max_concurrency: int | None = None,
 ) -> list[str]:
@@ -48,6 +55,9 @@ def build_acc_benchmark_cli_argv(
     ``--num-prompts`` / ``--max-concurrency`` defaults (overridable via ``ACC_BENCH_NUM_PROMPTS`` /
     ``ACC_BENCH_MAX_CONCURRENCY``), and when Daily-Omni runs adds ``--daily-omni-repo`` so Hub QA
     matches :func:`daily_omni_bench_argv` once ``run_acc_benchmark`` mirrors ``--daily-omni-repo`` into env.
+
+    Video-MME is **opt-in** (``skip_videomme=True`` by default) so existing Daily-Omni / Seed-TTS
+    callers stay unchanged; pass ``skip_videomme=False`` (emits ``--run-videomme``) to enable it.
     """
     n_prompts = int(os.environ.get("ACC_BENCH_NUM_PROMPTS", "2000")) if num_prompts is None else int(num_prompts)
     n_conc = int(os.environ.get("ACC_BENCH_MAX_CONCURRENCY", "10")) if max_concurrency is None else int(max_concurrency)
@@ -70,6 +80,12 @@ def build_acc_benchmark_cli_argv(
         argv.append("--skip-seed-tts")
     if skip_daily:
         argv.append("--skip-daily-omni")
+    if skip_videomme:
+        argv.append("--skip-videomme")
+    else:
+        argv.append("--run-videomme")
+        repo = os.environ.get("VLLM_VIDEOMME_REPO", DEFAULT_VIDEOMME_HF_REPO).strip() or DEFAULT_VIDEOMME_HF_REPO
+        argv.extend(["--videomme-repo", repo])
     return argv
 
 
@@ -99,6 +115,43 @@ def daily_omni_bench_argv() -> list[str]:
     ]
 
 
+def videomme_bench_argv() -> list[str]:
+    """CLI args for Video-MME (after ``vllm bench serve --omni``).
+
+    Same local-vs-Hub rule as :func:`seed_tts_bench_argv`:
+
+    * If ``VLLM_VIDEOMME_DATASET_PATH`` / ``VIDEOMME_ROOT`` names an existing directory,
+      use that local mirror as ``--dataset-path``.
+    * Otherwise pass the Hub id (``VLLM_VIDEOMME_REPO`` / ``lmms-eval/Video-MME``); the child
+      bench downloads via ``huggingface_hub.snapshot_download``.
+    """
+    root = os.environ.get("VLLM_VIDEOMME_DATASET_PATH", "").strip() or os.environ.get("VIDEOMME_ROOT", "").strip()
+    if root:
+        p = Path(root).expanduser()
+        # Preserve Hub ids verbatim. Only canonicalize when the value is a real directory.
+        if p.exists() and p.is_dir():
+            return [
+                "--dataset-name",
+                "videomme",
+                "--dataset-path",
+                str(p.resolve()),
+            ]
+        # Non-directory values (e.g. accidental Hub id in DATASET_PATH) fall through as path.
+        return [
+            "--dataset-name",
+            "videomme",
+            "--dataset-path",
+            root,
+        ]
+    repo = os.environ.get("VLLM_VIDEOMME_REPO", DEFAULT_VIDEOMME_HF_REPO).strip() or DEFAULT_VIDEOMME_HF_REPO
+    return [
+        "--dataset-name",
+        "videomme",
+        "--dataset-path",
+        repo,
+    ]
+
+
 def seed_tts_bench_argv(*, locale: str = "en") -> list[str]:
     """CLI args for Seed-TTS (after ``vllm bench serve --omni``)."""
     dp = os.environ.get("VLLM_SEED_TTS_DATASET_PATH", "").strip()
@@ -121,9 +174,13 @@ def seed_tts_bench_argv(*, locale: str = "en") -> list[str]:
 
 def find_vllm_cli() -> str:
     exe = shutil.which("vllm")
-    if not exe:
-        raise FileNotFoundError("Could not find `vllm` on PATH (install vLLM-Omni with CLI entrypoints).")
-    return exe
+    if exe:
+        return exe
+    # Fallback when pytest is invoked via absolute venv path without activating PATH.
+    sibling = Path(sys.executable).resolve().parent / "vllm"
+    if sibling.is_file() and os.access(sibling, os.X_OK):
+        return str(sibling)
+    raise FileNotFoundError("Could not find `vllm` on PATH (install vLLM-Omni with CLI entrypoints).")
 
 
 def run_vllm_bench_subprocess(vllm: str, argv: list[str], *, extra_env: dict[str, str] | None = None) -> None:
@@ -150,6 +207,11 @@ def build_serve_common_argv(
     result_dir: Path,
     result_filename: str,
     ready_check_timeout_sec: int | None = None,
+    trust_remote_code: bool = False,
+    backend: str = "openai-chat-omni",
+    endpoint: str = "/v1/chat/completions",
+    temperature: float | None = None,
+    output_len: int | None = None,
 ) -> list[str]:
     out = [
         "bench",
@@ -162,9 +224,9 @@ def build_serve_common_argv(
         "--model",
         model,
         "--endpoint",
-        "/v1/chat/completions",
+        endpoint,
         "--backend",
-        "openai-chat-omni",
+        backend,
         "--request-rate",
         "inf",
         "--num-prompts",
@@ -184,6 +246,12 @@ def build_serve_common_argv(
     ]
     if ready_check_timeout_sec is not None:
         out.extend(["--ready-check-timeout-sec", str(int(ready_check_timeout_sec))])
+    if trust_remote_code:
+        out.append("--trust-remote-code")
+    if temperature is not None:
+        out.extend(["--temperature", str(temperature)])
+    if output_len is not None:
+        out.extend(["--output-len", str(int(output_len))])
     return out
 
 
@@ -193,6 +261,12 @@ def assert_daily_omni_scored(result: dict[str, Any]) -> None:
     assert int(result.get("daily_omni_evaluated_ok", 0) or 0) > 0, "no successful MCQ rows (daily_omni_evaluated_ok==0)"
 
 
+def assert_videomme_scored(result: dict[str, Any]) -> None:
+    acc = result.get("videomme_accuracy")
+    assert acc is not None, "videomme_accuracy missing — wrong dataset or benchmark wiring"
+    assert int(result.get("videomme_evaluated_ok", 0) or 0) > 0, "no successful MCQ rows (videomme_evaluated_ok==0)"
+
+
 def assert_seed_tts_scored(result: dict[str, Any]) -> None:
     err = result.get("seed_tts_eval_setup_error")
     assert not err, f"Seed-TTS eval deps/setup failed: {err}"

tests/e2e/accuracy/qwen3_omni/run_qwen_omni_acc_benchmark.py

@@ -11,6 +11,9 @@ H100 / MI325) because they launch the live Omni server inside the test.
    run **fails** if accuracy is strictly below **0.69** (``--min-daily-omni-accuracy`` / ``ACC_BENCH_MIN_DAILY_OMNI_ACCURACY``).
 2. **Seed-TTS** — ``seed-tts-eval``-style metrics when ``--seed-tts-wer-eval`` is used
    (WER / SIM / UTMOS keys from :func:`compute_seed_tts_wer_metrics`).
+3. **Video-MME** — opt-in via ``--run-videomme`` (skipped by default so Daily-Omni / Seed-TTS callers
+   stay unchanged). MiniCPM-o 4.5 reports **70.4** on Video-MME (w/o subs); the default floor is
+   **0.68** (``--min-videomme-accuracy`` / ``ACC_BENCH_MIN_VIDEOMME_ACCURACY``).
 
 Prerequisites
 -------------
@@ -21,7 +24,8 @@ Prerequisites
   On L4 you may need a smaller checkpoint, quantization, or tighter engine flags; this script
   only drives the **client** benchmark.
 
-* ``vllm`` CLI from **vLLM-Omni** (so ``bench serve`` registers ``daily-omni`` / ``seed-tts``).
+* ``vllm`` CLI from **vLLM-Omni** (so ``bench serve`` registers ``daily-omni`` / ``seed-tts`` /
+  ``videomme``).
 
 * **Daily-Omni** — if local ``qa.json`` + ``Videos/`` are not both provided (CLI or matching env),
   the client passes ``--dataset-path`` with a Hub id (default ``liarliar/Daily-Omni``). The **child**
@@ -31,6 +35,13 @@ Prerequisites
   the dataset repo with ``--daily-omni-repo`` or ``VLLM_DAILY_OMNI_REPO``; override the tar repo
   with ``VLLM_DAILY_OMNI_MEDIA_REPO`` if needed.
 
+* **Video-MME** — by default ``--dataset-path`` is the Hub id ``lmms-eval/Video-MME`` (override with
+  ``--videomme-repo`` / ``VLLM_VIDEOMME_REPO``). The child bench downloads parquet + video archives via
+  ``huggingface_hub.snapshot_download`` (needs ``huggingface_hub``, network or HF cache). Pass an
+  existing local root with ``--videomme-dataset-path`` / ``VLLM_VIDEOMME_DATASET_PATH`` /
+  ``VIDEOMME_ROOT`` to skip the Hub download. For CI servers without ``--allowed-local-media-path``,
+  pass ``--videomme-inline-local-video`` (or rely on the pytest fixture that appends it).
+
 * **Seed-TTS** optional extras for WER/SIM/UTMOS::
 
     pip install 'vllm-omni[dev]'
@@ -55,6 +66,18 @@ Skip one suite, tighten gates::
         --skip-daily-omni \\
         --max-seed-tts-mean-wer 0.35 \\
         --min-seed-tts-mean-sim 0.75
+
+MiniCPM-o Video-MME (w/o subs)::
+
+    python tests/e2e/accuracy/qwen3_omni/run_qwen_omni_acc_benchmark.py \\
+        --skip-daily-omni --skip-seed-tts --run-videomme \\
+        --videomme-pack-mode minicpm-frames \\
+        --videomme-max-frames 96 \\
+        --temperature 0 --output-len 128 \\
+        --min-videomme-accuracy 0.68
+
+    # Optional local mirror (absolute path) instead of Hub download:
+    #   --videomme-dataset-path /path/to/Video-MME
 """
 
 from __future__ import annotations
@@ -75,6 +98,7 @@ from tests.e2e.accuracy.qwen3_omni.qwen3_omni_acc_bench_core import (
     load_benchmark_result,
     run_vllm_bench_subprocess,
     seed_tts_bench_argv,
+    videomme_bench_argv,
 )
 
 _REPO_ROOT = Path(__file__).resolve().parents[4]
@@ -102,6 +126,20 @@ def _validate_daily_omni(result: dict[str, Any], *, min_accuracy: float | None)
     return errs
 
 
+def _validate_videomme(result: dict[str, Any], *, min_accuracy: float | None) -> list[str]:
+    errs: list[str] = []
+    acc = result.get("videomme_accuracy")
+    if acc is None:
+        errs.append("Missing videomme_accuracy (wrong dataset or no gold-evaluated rows).")
+        return errs
+    ev = int(result.get("videomme_evaluated_ok", 0) or 0)
+    if ev <= 0:
+        errs.append("videomme_evaluated_ok is 0; no successful MCQ rows to score.")
+    if min_accuracy is not None and float(acc) + 1e-12 < float(min_accuracy):
+        errs.append(f"videomme_accuracy={acc:.6f} < --min-videomme-accuracy={min_accuracy}")
+    return errs
+
+
 def _validate_seed_tts(
     result: dict[str, Any],
     *,
@@ -130,7 +168,7 @@ def _validate_seed_tts(
 
 
 def sync_dataset_env_from_ns(ns: argparse.Namespace) -> None:
-    """Mirror CLI path flags into env vars read by ``daily_omni_bench_argv`` / ``seed_tts_bench_argv``."""
+    """Mirror CLI path flags into env vars read by dataset ``*_bench_argv`` helpers."""
     repo = getattr(ns, "daily_omni_repo", None)
     if repo is not None and str(repo).strip():
         os.environ["VLLM_DAILY_OMNI_REPO"] = str(repo).strip()
@@ -138,6 +176,14 @@ def sync_dataset_env_from_ns(ns: argparse.Namespace) -> None:
         os.environ["VLLM_DAILY_OMNI_QA_JSON"] = str(Path(ns.daily_omni_qa_json).expanduser().resolve())
     if ns.daily_omni_video_dir is not None:
         os.environ["VLLM_DAILY_OMNI_VIDEO_DIR"] = str(Path(ns.daily_omni_video_dir).expanduser().resolve())
+    vm_repo = getattr(ns, "videomme_repo", None)
+    if vm_repo is not None and str(vm_repo).strip():
+        os.environ["VLLM_VIDEOMME_REPO"] = str(vm_repo).strip()
+    vm_path = getattr(ns, "videomme_dataset_path", None)
+    if vm_path is not None and str(vm_path).strip():
+        raw = str(vm_path).strip()
+        p = Path(raw).expanduser()
+        os.environ["VLLM_VIDEOMME_DATASET_PATH"] = str(p.resolve()) if p.exists() and p.is_dir() else raw
     if ns.seed_tts_dataset_path is not None:
         # ``--seed-tts-dataset-path`` accepts either a local directory or a
         # Hugging Face repo id. Only resolve to an absolute filesystem path
@@ -157,6 +203,9 @@ def _preserve_benchmark_dataset_env() -> Any:
         "VLLM_DAILY_OMNI_REPO",
         "VLLM_DAILY_OMNI_QA_JSON",
         "VLLM_DAILY_OMNI_VIDEO_DIR",
+        "VLLM_VIDEOMME_REPO",
+        "VLLM_VIDEOMME_DATASET_PATH",
+        "VIDEOMME_ROOT",
         "VLLM_SEED_TTS_DATASET_PATH",
         "SEED_TTS_ROOT",
     )
@@ -171,7 +220,13 @@ def _preserve_benchmark_dataset_env() -> Any:
                 os.environ[key] = value
 
 
-def _build_common_args(ns: argparse.Namespace, *, result_filename: str) -> list[str]:
+def _build_common_args(
+    ns: argparse.Namespace,
+    *,
+    result_filename: str,
+    backend: str = "openai-chat-omni",
+    endpoint: str = "/v1/chat/completions",
+) -> list[str]:
     return build_serve_common_argv(
         host=ns.host,
         port=ns.port,
@@ -183,6 +238,11 @@ def _build_common_args(ns: argparse.Namespace, *, result_filename: str) -> list[
         result_dir=ns.result_dir,
         result_filename=result_filename,
         ready_check_timeout_sec=ns.ready_check_timeout_sec,
+        trust_remote_code=ns.trust_remote_code,
+        backend=backend,
+        endpoint=endpoint,
+        temperature=getattr(ns, "temperature", None),
+        output_len=getattr(ns, "output_len", None),
     )
 
 
@@ -201,6 +261,9 @@ def run_daily_omni(ns: argparse.Namespace, vllm: str) -> Path:
             json.dumps(extra, ensure_ascii=False, separators=(",", ":")),
         ]
     )
+    pack_mode = getattr(ns, "daily_omni_pack_mode", None)
+    if pack_mode:
+        argv.extend(["--daily-omni-pack-mode", pack_mode])
     if ns.daily_omni_save_eval_items:
         argv.append("--daily-omni-save-eval-items")
     print("\n$", vllm, *argv, "\n", flush=True)
@@ -211,13 +274,58 @@ def run_daily_omni(ns: argparse.Namespace, vllm: str) -> Path:
     return out
 
 
+def run_videomme(ns: argparse.Namespace, vllm: str) -> Path:
+    ns.result_dir.mkdir(parents=True, exist_ok=True)
+    tag = datetime.now().strftime("%Y%m%d-%H%M%S")
+    result_filename = f"omni_acc_videomme_{tag}.json"
+    extra = json.loads(ns.videomme_extra_body_json)
+    argv = (
+        _build_common_args(ns, result_filename=result_filename)
+        + videomme_bench_argv()
+        + [
+            "--disable-shuffle",
+            "--videomme-pack-mode",
+            ns.videomme_pack_mode,
+            "--videomme-duration",
+            ns.videomme_duration,
+            "--extra-body",
+            json.dumps(extra, ensure_ascii=False, separators=(",", ":")),
+        ]
+    )
+    if ns.videomme_max_frames is not None:
+        argv.extend(["--videomme-max-frames", str(int(ns.videomme_max_frames))])
+    if ns.videomme_use_subtitle:
+        argv.append("--videomme-use-subtitle")
+    if ns.videomme_inline_local_video:
+        argv.append("--videomme-inline-local-video")
+    if ns.videomme_save_eval_items:
+        argv.append("--videomme-save-eval-items")
+    if ns.videomme_video_dir is not None:
+        argv.extend(["--videomme-video-dir", str(Path(ns.videomme_video_dir).expanduser().resolve())])
+    if ns.videomme_parquet is not None:
+        argv.extend(["--videomme-parquet", str(Path(ns.videomme_parquet).expanduser().resolve())])
+    if ns.videomme_subtitle_dir is not None:
+        argv.extend(["--videomme-subtitle-dir", str(Path(ns.videomme_subtitle_dir).expanduser().resolve())])
+    print("\n$", vllm, *argv, "\n", flush=True)
+    run_vllm_bench_subprocess(vllm, argv)
+    out = Path(ns.result_dir) / result_filename
+    if not out.is_file():
+        raise FileNotFoundError(f"Expected result JSON at {out}")
+    return out
+
+
 def run_seed_tts(ns: argparse.Namespace, vllm: str) -> Path:
     ns.result_dir.mkdir(parents=True, exist_ok=True)
     tag = datetime.now().strftime("%Y%m%d-%H%M%S")
     result_filename = f"qwen_omni_acc_seed_tts_{tag}.json"
     extra = json.loads(ns.seed_extra_body_json)
     argv = (
-        _build_common_args(ns, result_filename=result_filename)
+        _build_common_args(
+            ns,
+            result_filename=result_filename,
+            backend=ns.seed_backend,
+            endpoint=ns.seed_endpoint,
+        )
         + seed_tts_bench_argv(locale=ns.seed_tts_locale)
         + [
             "--seed-tts-wer-eval",
@@ -225,6 +333,8 @@ def run_seed_tts(ns: argparse.Namespace, vllm: str) -> Path:
             json.dumps(extra, ensure_ascii=False, separators=(",", ":")),
         ]
     )
+    if ns.seed_tts_turns_per_session > 1:
+        argv.extend(["--seed-tts-turns-per-session", str(ns.seed_tts_turns_per_session)])
     if ns.seed_tts_wer_save_items:
         argv.append("--seed-tts-wer-save-items")
     if ns.seed_tts_file_ref_audio:
@@ -266,6 +376,23 @@ def build_arg_parser() -> argparse.ArgumentParser:
         help="If set, forwarded to ``vllm bench serve`` (probe first request until success). "
         "Omit to use upstream default (typically skip).",
     )
+    p.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Forward trust_remote_code to ``vllm bench serve`` for models with custom code.",
+    )
+    p.add_argument(
+        "--temperature",
+        type=float,
+        default=None,
+        help="Optional sampling temperature forwarded to ``vllm bench serve`` (e.g. 0 for MiniCPM-o Daily-Omni MCQ).",
+    )
+    p.add_argument(
+        "--output-len",
+        type=int,
+        default=None,
+        help="Optional max output tokens forwarded to ``vllm bench serve`` (e.g. 512 for MiniCPM-o Daily-Omni).",
+    )
     p.add_argument(
         "--result-dir",
         type=Path,
@@ -274,6 +401,20 @@ def build_arg_parser() -> argparse.ArgumentParser:
 
     p.add_argument("--skip-daily-omni", action="store_true")
     p.add_argument("--skip-seed-tts", action="store_true")
+    # Video-MME is opt-in: default skip=True so existing Daily-Omni / Seed-TTS callers stay unchanged.
+    p.set_defaults(skip_videomme=True)
+    p.add_argument(
+        "--skip-videomme",
+        action="store_true",
+        dest="skip_videomme",
+        help="Skip Video-MME (default).",
+    )
+    p.add_argument(
+        "--run-videomme",
+        action="store_false",
+        dest="skip_videomme",
+        help="Enable the Video-MME accuracy suite.",
+    )
 
     p.add_argument(
         "--daily-omni-repo",
@@ -295,6 +436,13 @@ def build_arg_parser() -> argparse.ArgumentParser:
         help="Optional local Videos root; if omitted, media is fetched lazily from Hub Videos.tar.",
     )
     p.add_argument("--daily-omni-input-mode", choices=("all", "visual", "audio"), default="all")
+    p.add_argument(
+        "--daily-omni-pack-mode",
+        choices=("qwen", "minicpm-interleave"),
+        default=None,
+        help="Daily-Omni multimodal pack recipe. Use ``minicpm-interleave`` for MiniCPM-o "
+        "(1fps interleaved image/audio; see PR #5606).",
+    )
     p.add_argument(
         "--daily-extra-body-json",
         default='{"modalities":["text"]}',
@@ -313,6 +461,87 @@ def build_arg_parser() -> argparse.ArgumentParser:
         "Default baseline 0.69; override with env ACC_BENCH_MIN_DAILY_OMNI_ACCURACY or pass 0 to disable the floor.",
     )
 
+    p.add_argument(
+        "--videomme-repo",
+        type=str,
+        default=None,
+        help="Hugging Face dataset id for Video-MME Hub mode (sets VLLM_VIDEOMME_REPO). "
+        "Default when no local path is set: lmms-eval/Video-MME. "
+        "Ignored when a local --videomme-dataset-path / VLLM_VIDEOMME_DATASET_PATH is used.",
+    )
+    p.add_argument(
+        "--videomme-dataset-path",
+        type=str,
+        default=None,
+        help="Optional local Video-MME root (parquet + videos). Existing directories skip the Hub "
+        "download; Hub ids are also accepted and passed through as --dataset-path. "
+        "Sets VLLM_VIDEOMME_DATASET_PATH.",
+    )
+    p.add_argument(
+        "--videomme-parquet",
+        type=Path,
+        default=None,
+        help="Optional explicit path to videomme/test-*.parquet.",
+    )
+    p.add_argument(
+        "--videomme-video-dir",
+        type=Path,
+        default=None,
+        help="Optional explicit directory of extracted Video-MME videos.",
+    )
+    p.add_argument(
+        "--videomme-subtitle-dir",
+        type=Path,
+        default=None,
+        help="Optional subtitle directory (used with --videomme-use-subtitle).",
+    )
+    p.add_argument(
+        "--videomme-pack-mode",
+        choices=("minicpm-frames", "minicpm-interleave", "video_url"),
+        default="minicpm-frames",
+        help="Video-MME multimodal pack recipe (OmniEvalKit MiniCPM default: minicpm-frames).",
+    )
+    p.add_argument(
+        "--videomme-max-frames",
+        type=int,
+        default=None,
+        help="Override max sampled frames (OmniEvalKit: 96 for minicpm-frames, 64 for interleave).",
+    )
+    p.add_argument(
+        "--videomme-duration",
+        choices=("all", "short", "medium", "long"),
+        default="all",
+        help="Filter Video-MME by duration bucket (default: all).",
+    )
+    p.add_argument(
+        "--videomme-use-subtitle",
+        action="store_true",
+        help="Prepend subtitle text to the user prompt (Video-MME w/ subs).",
+    )
+    p.add_argument(
+        "--videomme-inline-local-video",
+        action="store_true",
+        help="Embed frames/audio as base64 data URLs (no --allowed-local-media-path on the server).",
+    )
+    p.add_argument(
+        "--videomme-extra-body-json",
+        default='{"modalities":["text"],"chat_template_kwargs":{"enable_thinking":false}}',
+        help="JSON merged into each chat request for Video-MME.",
+    )
+    p.add_argument(
+        "--videomme-save-eval-items",
+        action="store_true",
+        help="Include per-request Video-MME accuracy rows in the saved JSON.",
+    )
+    p.add_argument(
+        "--min-videomme-accuracy",
+        type=float,
+        default=float((os.environ.get("ACC_BENCH_MIN_VIDEOMME_ACCURACY") or "0.68").strip() or "0.68"),
+        help="Fail when videomme_accuracy is strictly below this threshold (0–1). "
+        "Default 0.68 (~2pp under MiniCPM-o 4.5 reported 70.4 w/o subs); "
+        "override with ACC_BENCH_MIN_VIDEOMME_ACCURACY or pass 0 to disable the floor.",
+    )
+
     p.add_argument(
         "--seed-tts-dataset-path",
         type=str,
@@ -326,6 +555,22 @@ def build_arg_parser() -> argparse.ArgumentParser:
         default='{"modalities":["text","audio"]}',
         help="JSON for Seed-TTS chat requests (must include audio for synthesis + PCM capture).",
     )
+    p.add_argument(
+        "--seed-backend",
+        default="openai-chat-omni",
+        help="Seed-TTS benchmark backend (for example openai-realtime-tts).",
+    )
+    p.add_argument(
+        "--seed-endpoint",
+        default="/v1/chat/completions",
+        help="Seed-TTS serving endpoint corresponding to --seed-backend.",
+    )
+    p.add_argument(
+        "--seed-tts-turns-per-session",
+        type=int,
+        default=1,
+        help="Group this many Seed-TTS target texts into each Realtime session.",
+    )
     p.add_argument("--seed-tts-wer-save-items", action="store_true")
     p.add_argument(
         "--seed-tts-file-ref-audio",
@@ -366,7 +611,7 @@ def parse_acc_benchmark_args(argv: list[str] | None = None) -> argparse.Namespac
 
 
 def run_acc_benchmark(ns: argparse.Namespace) -> int:
-    """Run Daily-Omni and/or Seed-TTS client benches against a running server; return 0 on success."""
+    """Run configured client benches against a running server; return 0 on success."""
     failed: list[str] = []
 
     with _preserve_benchmark_dataset_env():
@@ -390,6 +635,21 @@ def run_acc_benchmark(ns: argparse.Namespace) -> int:
                     flush=True,
                 )
 
+        if not ns.skip_videomme:
+            path = run_videomme(ns, vllm)
+            print(f"\n[Video-MME] result JSON: {path}", flush=True)
+            data = load_benchmark_result(path)
+            errs = _validate_videomme(data, min_accuracy=ns.min_videomme_accuracy)
+            if errs:
+                failed.extend([f"[Video-MME] {e}" for e in errs])
+            else:
+                print(
+                    f"[Video-MME] videomme_accuracy={data.get('videomme_accuracy')} "
+                    f"evaluated_ok={data.get('videomme_evaluated_ok')} "
+                    f"per_duration={data.get('videomme_per_duration_accuracy')}",
+                    flush=True,
+                )
+
         if not ns.skip_seed_tts:
             path = run_seed_tts(ns, vllm)
             print(f"\n[Seed-TTS] result JSON: {path}", flush=True)
@@ -400,6 +660,11 @@ def run_acc_benchmark(ns: argparse.Namespace) -> int:
                 min_mean_sim=ns.min_seed_tts_mean_sim,
                 min_mean_utmos=ns.min_seed_tts_mean_utmos,
             )
+            if ns.seed_tts_turns_per_session > 1:
+                expected_turns = ns.num_prompts * ns.seed_tts_turns_per_session
+                evaluated_turns = int(data.get("seed_tts_content_evaluated") or 0)
+                if evaluated_turns != expected_turns:
+                    errs.append(f"seed_tts_content_evaluated={evaluated_turns} != expected {expected_turns}")
             if errs:
                 failed.extend([f"[Seed-TTS] {e}" for e in errs])
             else:

tests/e2e/offline_inference/test_minicpmo_4_5.py

@@ -14,7 +14,7 @@ from tests.helpers.stage_config import get_deploy_config_path
 
 models = ["openbmb/MiniCPM-o-4_5"]
 
-_CI_DEPLOY = get_deploy_config_path("minicpmo_4_5_batching.yaml")
+_CI_DEPLOY = get_deploy_config_path("minicpmo_4_5.yaml")
 
 
 test_params = [(model, None, {"deploy_config": _CI_DEPLOY, "trust_remote_code": True}) for model in models]
@@ -31,11 +31,10 @@ def get_question(prompt_type: str = "text") -> str:
     return prompts.get(prompt_type, prompts["text"])
 
 
-@pytest.mark.skip(reason="https://github.com/vllm-project/vllm-omni/issues/5437")
 @pytest.mark.core_model
 @pytest.mark.advanced_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_text_to_text(omni_runner, omni_runner_handler) -> None:
     """Test processing text, generating text output."""
@@ -45,7 +44,7 @@ def test_text_to_text(omni_runner, omni_runner_handler) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_audio_to_text(omni_runner, omni_runner_handler) -> None:
     """Test processing audio, generating text output."""
@@ -58,7 +57,7 @@ def test_audio_to_text(omni_runner, omni_runner_handler) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_image_to_text(omni_runner, omni_runner_handler) -> None:
     """Test processing image, generating text output."""
@@ -69,7 +68,7 @@ def test_image_to_text(omni_runner, omni_runner_handler) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_video_to_text(omni_runner, omni_runner_handler) -> None:
     """Test processing video, generating text output."""
@@ -80,19 +79,18 @@ def test_video_to_text(omni_runner, omni_runner_handler) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_text_to_audio(omni_runner, omni_runner_handler) -> None:
-    """Test processing text, generating audio output through the talker token2wav path."""
+    """Test processing text and generating audio through Talker and Code2Wav."""
     request_config = {"prompts": get_question("text"), "modalities": ["audio"]}
     omni_runner_handler.send_omni_request(request_config)
 
 
-@pytest.mark.skip(reason="https://github.com/vllm-project/vllm-omni/issues/5437")
 @pytest.mark.core_model
 @pytest.mark.advanced_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_mix_to_audio(omni_runner, omni_runner_handler) -> None:
     """Test processing mixed modalities (image + audio), generating audio output."""
@@ -111,7 +109,7 @@ def test_mix_to_audio(omni_runner, omni_runner_handler) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_video_to_audio(omni_runner, omni_runner_handler) -> None:
     """Test processing video, generating audio output."""

tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py

@@ -53,6 +53,7 @@ from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
     RealtimeEventCollector,
     build_realtime_url,
     read_pcm16_wav,
+    summarize_session_request_metrics,
 )
 
 _url_with_model = build_realtime_url
@@ -194,6 +195,38 @@ class DemoState:
             summaries[response_id] = timing
         return summaries
 
+    def session_request_metrics(
+        self,
+        *,
+        session_id: str | None = None,
+    ) -> list[dict[str, object]]:
+        """Return benchmark-style metrics for every response in this session."""
+        requests: list[dict[str, object]] = []
+        for request_index, (response_id, timing) in enumerate(self.response_timing_summaries().items()):
+            metrics = timing.get("request_metrics")
+            if not isinstance(metrics, dict):
+                continue
+            requests.append(
+                {
+                    "session_id": session_id,
+                    "request_index": request_index,
+                    "response_id": response_id,
+                    **metrics,
+                }
+            )
+        return requests
+
+    def session_metric_summary(
+        self,
+        *,
+        session_id: str | None = None,
+    ) -> dict[str, object]:
+        """Average TTFT, TTFP, and RTF across responses that emitted audio."""
+        return summarize_session_request_metrics(
+            self.session_request_metrics(session_id=session_id),
+            session_id=session_id,
+        )
+
     def first_index(self, event_type: str, predicate=None) -> int | None:
         for index, event in enumerate(self.events):
             if event.get("type") != event_type:
@@ -930,12 +963,23 @@ async def _send_clean_turn(
         if outcome == "listen":
             return None, "listen"
     else:
-        await _wait_for(
-            state,
-            lambda: state.count("response.created") > before_created,
-            timeout_s=timeout_s,
-            label=f"{transcript} response.created",
-        )
+        try:
+            await _wait_for(
+                state,
+                lambda: state.count("response.created") > before_created,
+                timeout_s=timeout_s,
+                label=f"{transcript} response.created",
+            )
+        except TimeoutError:
+            # ``auto_response`` sessions only get ``response.created`` once the model
+            # decides to speak, so report the listen decisions that took its place.
+            raise TimeoutError(
+                f"Timed out waiting for {transcript} response.created after "
+                f"{state.count('input_audio_buffer.committed')} committed input turns; "
+                f"model_listen={state.model_listen_count - before_model_listen} "
+                f"buffering_listen={state.buffering_listen_count} "
+                f"errors={state.count('error')}"
+            ) from None
         await _wait_for(
             state,
             lambda: eligible_response_id() is not None,
@@ -1174,10 +1218,13 @@ async def _send_listen_only_overlap_pair(
     before_second_listen = state.model_listen_count
     overlap_started_while_active = not state.response_done(first_response_id)
     model_unit_ready_while_active = False
+    second_unit_event_index: int | None = None
 
     def record_model_unit_ready() -> None:
-        nonlocal model_unit_ready_while_active
+        nonlocal model_unit_ready_while_active, second_unit_event_index
         model_unit_ready_while_active = not state.response_done(first_response_id)
+        if model_unit_ready_while_active:
+            second_unit_event_index = len(state.events)
 
     await _send_pcm16(
         ws,
@@ -1239,14 +1286,13 @@ async def _send_listen_only_overlap_pair(
             label=transcripts[1],
         )
 
-    second_speech_index = _nth_event_index(state, "input_audio_buffer.speech_started", 2)
     first_done_index = _event_index_for_response(state, "response.done", first_response_id)
     overlap_ok = (
         overlap_started_while_active
         and model_unit_ready_while_active
-        and second_speech_index is not None
+        and second_unit_event_index is not None
         and first_done_index is not None
-        and second_speech_index < first_done_index
+        and second_unit_event_index < first_done_index
     )
     return [first_response_id, second_response_id], ["speak", second_outcome], overlap_ok
 
@@ -1380,6 +1426,17 @@ async def run_demo(args: argparse.Namespace) -> dict[str, object]:
             await ws.send(json.dumps({"type": "session.close"}))
             await _wait_for(state, lambda: state.count("session.closed") > 0, timeout_s=20, label="session.closed")
         finally:
+            if state.count("session.created") > 0 and state.count("session.closed") == 0:
+                try:
+                    await ws.send(json.dumps({"type": "session.close"}))
+                    await _wait_for(
+                        state,
+                        lambda: state.count("session.closed") > 0,
+                        timeout_s=min(args.timeout_s, 5),
+                        label="session.closed during cleanup",
+                    )
+                except (ConnectionClosed, TimeoutError):
+                    pass
             stop.set()
             reader.cancel()
             try:
@@ -1510,6 +1567,8 @@ async def run_demo(args: argparse.Namespace) -> dict[str, object]:
         "input_transcription_ok": input_transcription_ok,
         "completed_response_ids": completed_response_ids,
         "response_timings": state.response_timing_summaries(),
+        "request_metrics": state.session_request_metrics(session_id=args.session_id),
+        "session_metrics": state.session_metric_summary(session_id=args.session_id),
         "lifecycle_counts_ok": lifecycle_counts_ok,
         "validation_mode": validation_mode,
         "scenario": scenario,

tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py

@@ -8,6 +8,7 @@ import hashlib
 import json
 import sys
 import uuid
+from functools import lru_cache
 from pathlib import Path
 from types import SimpleNamespace
 from urllib.parse import parse_qsl, urlencode, urlsplit, urlunsplit
@@ -19,7 +20,25 @@ SCRIPT_DIR = Path(__file__).resolve().parent
 if str(SCRIPT_DIR) not in sys.path:
     sys.path.insert(0, str(SCRIPT_DIR))
 
-from minicpmo_realtime_duplex_scenarios import _url_with_model, run_demo  # noqa: E402
+
+@lru_cache(maxsize=1)
+def _scenario_module():
+    """Delay heavyweight vLLM imports so ``--help`` remains directly executable."""
+    import minicpmo_realtime_duplex_scenarios
+
+    return minicpmo_realtime_duplex_scenarios
+
+
+def _ref_audio_data_url(path: str) -> str:
+    return _scenario_module()._ref_audio_data_url(path)
+
+
+def _url_with_model(*args, **kwargs) -> str:
+    return _scenario_module()._url_with_model(*args, **kwargs)
+
+
+async def run_demo(args):
+    return await _scenario_module().run_demo(args)
 
 
 def _with_resume_mode(url: str) -> str:
@@ -114,7 +133,12 @@ async def _open_admission_session(
     args: argparse.Namespace,
     session_id: str,
 ) -> tuple[object, dict[str, object]]:
-    url = _url_with_model(args.url, args.model, session_id=session_id)
+    url = _url_with_model(
+        args.url,
+        args.model,
+        autostart=False if getattr(args, "ref_audio", None) else None,
+        session_id=session_id,
+    )
     ws = await websockets.connect(url, max_size=64 * 1024 * 1024)
     await ws.send(
         json.dumps(
@@ -125,6 +149,7 @@ async def _open_admission_session(
                     "model": args.model,
                     "modalities": ["audio", "text"],
                     "extra_body": {"minicpmo45_native_duplex": True},
+                    **({"ref_audio": _ref_audio_data_url(args.ref_audio)} if getattr(args, "ref_audio", None) else {}),
                 },
             }
         )
@@ -151,7 +176,12 @@ async def _admission_probe(args: argparse.Namespace, *, limit: int) -> dict[str,
             accepted.append(await _open_admission_session(args, f"{prefix}-accepted-{index}"))
 
         overflow_id = f"{prefix}-overflow"
-        overflow_url = _url_with_model(args.url, args.model, session_id=overflow_id)
+        overflow_url = _url_with_model(
+            args.url,
+            args.model,
+            autostart=False if getattr(args, "ref_audio", None) else None,
+            session_id=overflow_id,
+        )
         async with websockets.connect(overflow_url, max_size=64 * 1024 * 1024) as overflow:
             await overflow.send(
                 json.dumps(
@@ -162,6 +192,11 @@ async def _admission_probe(args: argparse.Namespace, *, limit: int) -> dict[str,
                             "model": args.model,
                             "modalities": ["audio", "text"],
                             "extra_body": {"minicpmo45_native_duplex": True},
+                            **(
+                                {"ref_audio": _ref_audio_data_url(args.ref_audio)}
+                                if getattr(args, "ref_audio", None)
+                                else {}
+                            ),
                         },
                     }
                 )
@@ -207,16 +242,27 @@ async def _resume_probe(
     session_id: str,
     expect_expired: bool = False,
 ) -> dict[str, object]:
-    url = _url_with_model(args.url, args.model, session_id=session_id)
+    url = _url_with_model(
+        args.url,
+        args.model,
+        autostart=False if getattr(args, "ref_audio", None) else None,
+        session_id=session_id,
+    )
     async with websockets.connect(url, max_size=64 * 1024 * 1024) as first:
         await first.send(
             json.dumps(
                 {
                     "type": "session.update",
                     "session": {
+                        "session_id": session_id,
                         "model": args.model,
                         "modalities": ["audio", "text"],
                         "extra_body": {"minicpmo45_native_duplex": True},
+                        **(
+                            {"ref_audio": _ref_audio_data_url(args.ref_audio)}
+                            if getattr(args, "ref_audio", None)
+                            else {}
+                        ),
                     },
                 }
             )
@@ -306,7 +352,12 @@ async def _takeover_probe(
     *,
     session_id: str,
 ) -> dict[str, object]:
-    url = _url_with_model(args.url, args.model, session_id=session_id)
+    url = _url_with_model(
+        args.url,
+        args.model,
+        autostart=False if getattr(args, "ref_audio", None) else None,
+        session_id=session_id,
+    )
     first = await websockets.connect(url, max_size=64 * 1024 * 1024)
     second = None
     try:
@@ -315,9 +366,15 @@ async def _takeover_probe(
                 {
                     "type": "session.update",
                     "session": {
+                        "session_id": session_id,
                         "model": args.model,
                         "modalities": ["audio", "text"],
                         "extra_body": {"minicpmo45_native_duplex": True},
+                        **(
+                            {"ref_audio": _ref_audio_data_url(args.ref_audio)}
+                            if getattr(args, "ref_audio", None)
+                            else {}
+                        ),
                     },
                 }
             )
@@ -451,20 +508,7 @@ async def run_multi_session(args: argparse.Namespace) -> dict[str, object]:
             args,
             session_id=f"takeover-{args.takeover_session_index}-{uuid.uuid4().hex}",
         )
-    expiry_result = None
-    if args.expire_session_index is not None:
-        if not 0 <= args.expire_session_index < args.sessions:
-            raise ValueError("--expire-session-index is outside the session range")
-        expiry_result = await _resume_probe(
-            args,
-            session_id=f"expire-{args.expire_session_index}-{uuid.uuid4().hex}",
-            expect_expired=True,
-        )
-    admission_result = (
-        await _admission_probe(args, limit=args.verify_admission_limit)
-        if args.verify_admission_limit is not None
-        else None
-    )
+    lifecycle_result = await run_lifecycle_probes(args)
 
     session_results = await asyncio.gather(
         *(run_demo(_demo_args(args, index)) for index in range(args.sessions)),
@@ -487,16 +531,15 @@ async def run_multi_session(args: argparse.Namespace) -> dict[str, object]:
             and semantic_isolation_ok
             and (resume_result is None or resume_result.get("ok") is True)
             and (takeover_result is None or takeover_result.get("ok") is True)
-            and (expiry_result is None or expiry_result.get("ok") is True)
-            and (admission_result is None or admission_result.get("ok") is True)
+            and lifecycle_result["ok"] is True
         ),
         "session_count": args.sessions,
         "identity_isolation_ok": identity_isolation_ok,
         "semantic_isolation_ok": semantic_isolation_ok,
         "resume": resume_result,
         "takeover": takeover_result,
-        "expiry": expiry_result,
-        "admission": admission_result,
+        "expiry": lifecycle_result["expiry"],
+        "admission": lifecycle_result["admission"],
         "failures": failures,
         "sessions": completed,
     }
@@ -507,6 +550,34 @@ async def run_multi_session(args: argparse.Namespace) -> dict[str, object]:
     return result
 
 
+async def run_lifecycle_probes(args: argparse.Namespace) -> dict[str, object]:
+    """Run expiry and admission probes without requiring model output."""
+    if args.sessions < 1:
+        raise ValueError("--sessions must be positive")
+    expiry_result = None
+    if args.expire_session_index is not None:
+        if not 0 <= args.expire_session_index < args.sessions:
+            raise ValueError("--expire-session-index is outside the session range")
+        expiry_result = await _resume_probe(
+            args,
+            session_id=f"expire-{args.expire_session_index}-{uuid.uuid4().hex}",
+            expect_expired=True,
+        )
+    admission_result = (
+        await _admission_probe(args, limit=args.verify_admission_limit)
+        if args.verify_admission_limit is not None
+        else None
+    )
+    return {
+        "ok": (
+            (expiry_result is None or expiry_result.get("ok") is True)
+            and (admission_result is None or admission_result.get("ok") is True)
+        ),
+        "expiry": expiry_result,
+        "admission": admission_result,
+    }
+
+
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser()
     parser.add_argument("--url", default="ws://127.0.0.1:8113/v1/realtime?duplex=1")

tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py

@@ -2,8 +2,8 @@
 
 This E2E driver runs the public ``realtime_duplex_demo.py`` against a live
 duplex backend. Arbitrary audio defaults to the model-policy lifecycle contract.
-The stronger response-required mode binds the two-response contract to a known
-input checksum and expected second response.
+The stronger response-required mode binds the multi-response contract to a
+known input checksum and expected follow-up response.
 """
 
 from __future__ import annotations
@@ -39,6 +39,17 @@ def _validate_input_sha256(path: Path, expected: str) -> str:
     return actual
 
 
+def _input_duration_s(path: Path) -> float:
+    with wave.open(str(path), "rb") as wav_file:
+        return wav_file.getnframes() / wav_file.getframerate()
+
+
+def _client_process_timeout_s(input_wav: Path, protocol_timeout_s: float) -> float:
+    # The child first streams the WAV in real time, then can independently
+    # exhaust its post-commit and session-close protocol waits.
+    return _input_duration_s(input_wav) + 2 * protocol_timeout_s + 30.0
+
+
 def _read_json(path: Path) -> dict[str, object]:
     if not path.is_file():
         return {}
@@ -211,7 +222,7 @@ def summarize_artifacts(
     validation_mode: str,
     min_responses: int,
     min_audio_deltas_per_response: int,
-    expect_second_response_substring: str | None,
+    expect_followup_response_substring: str | None,
 ) -> dict[str, object]:
     if validation_mode not in {"model-policy", "response-required"}:
         raise ValueError(f"unsupported validation mode: {validation_mode}")
@@ -274,12 +285,20 @@ def summarize_artifacts(
     )
     final_listen_after_commit = commit_index is not None and any(index > commit_index for index in listen_indices)
     transcript = "".join(str(summary.get("transcript") or "") for summary in response_summaries)
-    second_response_transcript = (
-        str(response_summaries[1].get("transcript") or "") if len(response_summaries) >= 2 else ""
+    followup_response_transcripts = [
+        str(summary.get("transcript") or "")
+        for summary in response_summaries[1:]
+        if commit_index is not None
+        and isinstance(summary.get("created_index"), int)
+        and summary["created_index"] < commit_index
+    ]
+    followup_response_transcript_ok = any(
+        bool(_normalize_text(transcript)) for transcript in followup_response_transcripts
+    )
+    followup_response_transcript_expectation_ok = not expect_followup_response_substring or any(
+        _normalize_text(expect_followup_response_substring) in _normalize_text(transcript)
+        for transcript in followup_response_transcripts
     )
-    second_response_transcript_expectation_ok = not expect_second_response_substring or _normalize_text(
-        expect_second_response_substring
-    ) in _normalize_text(second_response_transcript)
     error_events = [
         event
         for event in events
@@ -310,9 +329,8 @@ def summarize_artifacts(
     mode_contract_ok = validation_mode == "model-policy" or (
         second_response_before_final_commit
         and listen_before_first_response
-        and listen_between_responses
         and listen_after_last_done
-        and second_response_transcript_expectation_ok
+        and followup_response_transcript_ok
     )
     ok = common_contract_ok and mode_contract_ok
     return {
@@ -337,9 +355,10 @@ def summarize_artifacts(
         "listen_after_response_before_commit": listen_after_response_before_commit,
         "final_listen_after_commit": final_listen_after_commit,
         "transcript": transcript,
-        "second_response_transcript": second_response_transcript,
-        "expect_second_response_substring": expect_second_response_substring,
-        "second_response_transcript_expectation_ok": second_response_transcript_expectation_ok,
+        "followup_response_transcripts": followup_response_transcripts,
+        "followup_response_transcript_ok": followup_response_transcript_ok,
+        "expect_followup_response_substring": expect_followup_response_substring,
+        "followup_response_transcript_expectation_ok": followup_response_transcript_expectation_ok,
         "error_count": len(error_events),
         "cancelled_count": cancelled_count,
         "compact_sequence": _compact_sequence(events),
@@ -388,7 +407,7 @@ async def run_soft_interrupt(args: argparse.Namespace) -> dict[str, object]:
     try:
         stdout_bytes, stderr_bytes = await asyncio.wait_for(
             process.communicate(),
-            timeout=args.timeout_s + 30.0,
+            timeout=_client_process_timeout_s(input_wav, args.timeout_s),
         )
     except TimeoutError:
         process.kill()
@@ -406,7 +425,7 @@ async def run_soft_interrupt(args: argparse.Namespace) -> dict[str, object]:
         validation_mode=args.validation_mode,
         min_responses=args.min_responses,
         min_audio_deltas_per_response=args.min_audio_deltas_per_response,
-        expect_second_response_substring=args.expect_second_response_substring,
+        expect_followup_response_substring=args.expect_followup_response_substring,
     )
     summary.update(
         {
@@ -444,14 +463,12 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--min-responses", type=int, default=2)
     parser.add_argument("--min-audio-deltas-per-response", type=int, default=2)
     parser.add_argument("--input-sha256")
-    parser.add_argument("--expect-second-response-substring")
+    parser.add_argument("--expect-followup-response-substring")
     args = parser.parse_args()
     if args.validation_mode == "response-required" and args.min_responses < 2:
         parser.error("--min-responses must be at least 2")
     if args.validation_mode == "response-required" and not args.input_sha256:
         parser.error("--input-sha256 is required in response-required mode")
-    if args.validation_mode == "response-required" and not args.expect_second_response_substring:
-        parser.error("--expect-second-response-substring is required in response-required mode")
     if args.min_audio_deltas_per_response < 1:
         parser.error("--min-audio-deltas-per-response must be positive")
     return args

tests/e2e/online_serving/test_minicpmo_4_5.py

@@ -1,9 +1,7 @@
-"""
-E2E Online tests for MiniCPM-o 4.5 model with multimodal input and audio / text output.
+"""E2E online tests for MiniCPM-o 4.5 multimodal input and audio/text output.
 
-MiniCPM-o 4.5 has ``async_chunk: false``, ``max_num_seqs: 1`` on both stages,
-and the vocoder runs in-process inside the talker stage rather than as a separate
-Code2Wav stage.
+Exercises async chunk streaming (``--async-chunk``) across separate Thinker,
+Talker, and Code2Wav stages.
 """
 
 import os
@@ -13,26 +11,28 @@ import pytest
 from tests.helpers.mark import hardware_test
 from tests.helpers.media import generate_synthetic_audio, generate_synthetic_image, generate_synthetic_video
 from tests.helpers.runtime import OmniServerParams, dummy_messages_from_mix_data
-from tests.helpers.stage_config import get_deploy_config_path
+from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
 _MODEL = "openbmb/MiniCPM-o-4_5"
-_CI_DEPLOY = get_deploy_config_path("minicpmo_4_5_batching.yaml")
+_CI_DEPLOY = modify_stage_config(
+    get_deploy_config_path("minicpmo_4_5.yaml"),
+    updates={
+        "stages": {0: {"default_sampling_params.max_tokens": 64}, 1: {"default_sampling_params.max_tokens": 1024}}
+    },
+)
 
 test_params = [
     pytest.param(
         OmniServerParams(
             model=_MODEL,
             stage_config_path=_CI_DEPLOY,
-            use_stage_cli=True,
-            server_args=[
-                "--trust-remote-code",
-                "--no-async-chunk",
-            ],
+            use_stage_cli=False,
+            server_args=["--trust-remote-code", "--async-chunk"],
         ),
-        id="default",
-    )
+        id="async_chunk",
+    ),
 ]
 
 
@@ -65,16 +65,15 @@ def get_max_batch_size(size_type="few"):
     return batch_sizes.get(size_type, 5)
 
 
-@pytest.mark.skip(reason="https://github.com/vllm-project/vllm-omni/issues/5437")
 @pytest.mark.core_model
 @pytest.mark.advanced_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_text_to_text_001(omni_server, openai_client) -> None:
     """
     Test text-only input generating text output via OpenAI API.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text
     Output Modal: text
     Input Setting: stream=False
@@ -94,13 +93,13 @@ def test_text_to_text_001(omni_server, openai_client) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_text_to_audio_001(omni_server, openai_client) -> None:
     """
     Test text-only input generating text + audio output via OpenAI API.
-    This exercises the talker TTS region detection and in-process token2wav vocoder.
-    Deploy Setting: default 2GPU
+    This exercises Talker TTS region detection and the Code2Wav stage.
+    Deploy Setting: default single GPU
     Input Modal: text
     Output Modal: text + audio
     Input Setting: stream=True
@@ -119,12 +118,12 @@ def test_text_to_audio_001(omni_server, openai_client) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_audio_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Test audio input generating text + audio output via OpenAI API.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text + audio
     Output Modal: text + audio
     Input Setting: stream=True
@@ -133,13 +132,21 @@ def test_audio_to_text_audio_001(omni_server, openai_client) -> None:
     messages = dummy_messages_from_mix_data(
         system_prompt=get_system_prompt(),
         audio_data_url=audio_data_url,
-        content_text=get_prompt("mix"),
+        content_text=get_prompt(),
     )
 
     request_config = {
         "model": omni_server.model,
         "messages": messages,
         "stream": True,
+        "key_words": {"text": ["Beijing"]},
+        "modalities": ["text", "audio"],
+        "extra_body": {
+            "chat_template_kwargs": {
+                "use_tts_template": True,
+                "enable_thinking": False,
+            }
+        },
     }
 
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
@@ -147,12 +154,12 @@ def test_audio_to_text_audio_001(omni_server, openai_client) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_image_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Test image input generating text + audio output via OpenAI API.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text + image
     Output Modal: text + audio
     Input Setting: stream=True
@@ -175,12 +182,12 @@ def test_image_to_text_audio_001(omni_server, openai_client) -> None:
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_video_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Test video input generating text + audio output via OpenAI API.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text + video
     Output Modal: text + audio
     Input Setting: stream=True
@@ -201,16 +208,15 @@ def test_video_to_text_audio_001(omni_server, openai_client) -> None:
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
 
 
-@pytest.mark.skip(reason="https://github.com/vllm-project/vllm-omni/issues/5437")
 @pytest.mark.core_model
 @pytest.mark.advanced_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_mix_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Test multi-modal input (text + audio + video + image) generating text + audio output.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text + audio + video + image
     Output Modal: text + audio
     Input Setting: stream=True

tests/e2e/online_serving/test_minicpmo_4_5_duplex.py

@@ -0,0 +1,176 @@
+"""CI coverage for the MiniCPM-o 4.5 native-duplex Realtime API."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import uuid
+from pathlib import Path
+
+import pytest
+import websockets
+
+from tests.e2e.online_serving.minicpmo_realtime_duplex_scenarios import (
+    _ref_audio_data_url,
+    run_demo,
+)
+from tests.e2e.online_serving.run_minicpmo_realtime_duplex_multi_session import (
+    run_multi_session,
+)
+from tests.helpers.mark import hardware_test
+from tests.helpers.minicpmo_4_5_duplex import (
+    SERVER_PARAMS,
+    demo_args,
+    multi_session_args,
+    realtime_url,
+    resolve_ref_audio,
+    validated_input_wav,
+)
+from vllm_omni.experimental.fullduplex.client import build_realtime_url
+
+pytestmark = pytest.mark.omni
+
+
+def _assert_request_metrics(metrics: object, *, expected_count: int) -> None:
+    assert isinstance(metrics, list)
+    assert len(metrics) == expected_count
+    for request_index, request in enumerate(metrics):
+        assert isinstance(request["session_id"], str)
+        assert request["request_index"] == request_index
+        assert isinstance(request["response_id"], str)
+        assert request["ttft_ms"] is not None and request["ttft_ms"] >= 0
+        assert request["ttfp_ms"] >= 0
+        assert request["rtf"] is not None and request["rtf"] >= 0
+        assert request["audio_generation_ms"] >= 0
+        assert request["audio_duration_ms"] > 0
+
+
+def _assert_session_metrics(metrics: object, *, expected_count: int) -> None:
+    assert isinstance(metrics, dict)
+    assert isinstance(metrics["session_id"], str)
+    assert metrics["audio_turn_count"] == expected_count
+    assert metrics["mean_ttft_ms"] is not None and metrics["mean_ttft_ms"] >= 0
+    assert metrics["mean_ttfp_ms"] is not None and metrics["mean_ttfp_ms"] >= 0
+    assert metrics["mean_rtf"] is not None and metrics["mean_rtf"] >= 0
+
+
+async def _receive_protocol_events(ws, required_types: set[str], *, timeout_s: float) -> list[dict[str, object]]:
+    async def receive() -> list[dict[str, object]]:
+        events: list[dict[str, object]] = []
+        seen: set[str] = set()
+        while not required_types.issubset(seen):
+            raw = await ws.recv()
+            if not isinstance(raw, str):
+                continue
+            event = json.loads(raw)
+            if not isinstance(event, dict):
+                continue
+            events.append(event)
+            event_type = event.get("type")
+            if event_type == "error":
+                raise AssertionError(f"WebSocket protocol smoke received an error: {event}")
+            if isinstance(event_type, str):
+                seen.add(event_type)
+        return events
+
+    return await asyncio.wait_for(receive(), timeout=timeout_s)
+
+
+async def _run_protocol_smoke(*, url: str, model: str, ref_audio: Path) -> list[dict[str, object]]:
+    session_id = f"duplex-ci-protocol-{uuid.uuid4().hex}"
+    websocket_url = build_realtime_url(url, model, autostart=False, session_id=session_id)
+    async with websockets.connect(websocket_url, max_size=64 * 1024 * 1024) as ws:
+        await ws.send(
+            json.dumps(
+                {
+                    "type": "session.update",
+                    "session": {
+                        "session_id": session_id,
+                        "model": model,
+                        "modalities": ["audio", "text"],
+                        "ref_audio": _ref_audio_data_url(str(ref_audio)),
+                        "extra_body": {"minicpmo45_native_duplex": True},
+                    },
+                }
+            )
+        )
+        events = await _receive_protocol_events(
+            ws,
+            {"session.created", "session.updated"},
+            timeout_s=60,
+        )
+        await ws.send(json.dumps({"type": "session.close"}))
+        events.extend(await _receive_protocol_events(ws, {"session.closed"}, timeout_s=60))
+    return events
+
+
+@pytest.mark.core_model
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_websocket_protocol_smoke(omni_server, model_prefix: str) -> None:
+    ref_audio = resolve_ref_audio(model_prefix)
+    events = asyncio.run(
+        _run_protocol_smoke(
+            url=realtime_url(omni_server),
+            model=omni_server.model,
+            ref_audio=ref_audio,
+        )
+    )
+    event_types = [event.get("type") for event in events]
+    assert "session.created" in event_types
+    assert "session.updated" in event_types
+    assert event_types[-1] == "session.closed"
+
+
+@pytest.mark.advanced_model
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_single_session_response_required(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    args = demo_args(
+        omni_server=omni_server,
+        input_wav=validated_input_wav(),
+        ref_audio=resolve_ref_audio(model_prefix),
+        output_dir=tmp_path / "single_session",
+    )
+    args.turns = 2
+    # Every turn replays the same active-speech window as the first one. The default
+    # shorter follow-up window is a different mid-utterance slice, which the native
+    # duplex model may legitimately answer with "listen" instead of a response.
+    args.turn_duration_ms = [args.first_turn_ms] * args.turns
+    result = asyncio.run(run_demo(args))
+    assert result["ok"] is True
+    assert result["audio_delta_count"] > 0
+    assert result["done_count"] == 2
+    assert result["error_count"] == 0
+    assert result["all_audio_responses_have_transcript"] is True
+    assert result["transcript_delta_done_ok"] is True
+    _assert_request_metrics(result["request_metrics"], expected_count=2)
+    _assert_session_metrics(result["session_metrics"], expected_count=2)
+
+
+@pytest.mark.advanced_model
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_two_sessions_resume_and_takeover(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    result = asyncio.run(
+        run_multi_session(
+            multi_session_args(
+                omni_server=omni_server,
+                input_wav=validated_input_wav(),
+                ref_audio=resolve_ref_audio(model_prefix),
+                output_dir=tmp_path / "multi_session",
+                response_required=True,
+            )
+        )
+    )
+    assert result["ok"] is True
+    assert result["session_count"] == 2
+    assert result["resume"]["ok"] is True
+    assert result["takeover"]["ok"] is True
+    assert not result["failures"]
+    assert all(session["audio_delta_count"] > 0 for session in result["sessions"])
+    assert all(session["done_count"] == 1 for session in result["sessions"])
+    assert all(session["error_count"] == 0 for session in result["sessions"])
+    for session in result["sessions"]:
+        _assert_request_metrics(session["request_metrics"], expected_count=1)
+        _assert_session_metrics(session["session_metrics"], expected_count=1)

tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py

@@ -0,0 +1,90 @@
+"""Nightly lifecycle coverage for the MiniCPM-o 4.5 native-duplex API."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from tests.e2e.online_serving.run_minicpmo_realtime_duplex_multi_session import (
+    run_lifecycle_probes,
+)
+from tests.e2e.online_serving.run_minicpmo_realtime_duplex_soft_interrupt import (
+    run_soft_interrupt,
+)
+from tests.helpers.mark import hardware_test
+from tests.helpers.minicpmo_4_5_duplex import (
+    SERVER_PARAMS,
+    SOFT_INTERRUPT_SHA256,
+    multi_session_args,
+    realtime_url,
+    resolve_ref_audio,
+    validated_input_wav,
+    validated_soft_interrupt_wav,
+)
+
+pytestmark = [pytest.mark.full_model, pytest.mark.omni]
+
+
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_admission_and_expiry_reaper(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    args = multi_session_args(
+        omni_server=omni_server,
+        input_wav=validated_input_wav(),
+        ref_audio=resolve_ref_audio(model_prefix),
+        output_dir=tmp_path / "admission_expiry",
+        response_required=False,
+    )
+    args.sessions = 1
+    args.disconnect_session_index = None
+    args.takeover_session_index = None
+    args.expire_session_index = 0
+    args.verify_admission_limit = 2
+    result = asyncio.run(run_lifecycle_probes(args))
+
+    assert result["ok"] is True, json.dumps(result, ensure_ascii=False, indent=2)
+    assert result["expiry"]["ok"] is True
+    assert result["expiry"]["error_code"] == "session_resume_expired"
+    assert result["admission"]["ok"] is True
+    assert result["admission"]["overflow_error_code"] == "resource_exhausted"
+
+
+# CUDA-only for now: this contract asserts the model speaks while the user is
+# still talking, which only holds when the duplex pipeline sustains real-time
+# throughput. The current NPU stack runs several times slower than real time,
+# so it never reaches a mid-stream decision point.
+@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_soft_interrupt(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    input_wav = validated_soft_interrupt_wav()
+    result = asyncio.run(
+        run_soft_interrupt(
+            SimpleNamespace(
+                url=realtime_url(omni_server),
+                model=omni_server.model,
+                input_wav=str(input_wav),
+                ref_audio=str(resolve_ref_audio(model_prefix)),
+                output_dir=str(tmp_path / "soft_interrupt"),
+                summary_output=None,
+                chunk_ms=200,
+                timeout_s=180.0,
+                require_audio=True,
+                no_realtime_pacing=False,
+                validation_mode="response-required",
+                min_responses=2,
+                min_audio_deltas_per_response=2,
+                input_sha256=SOFT_INTERRUPT_SHA256,
+                expect_followup_response_substring=None,
+            )
+        )
+    )
+
+    assert result["ok"] is True, json.dumps(result, ensure_ascii=False, indent=2)
+    assert result["error_count"] == 0
+    assert result["response_lifecycle_ok"] is True
+    assert result["response_audio_contract_ok"] is True
+    assert result["followup_response_transcript_ok"] is True

tests/e2e/online_serving/test_minicpmo_4_5_expansion.py

@@ -1,7 +1,10 @@
-"""
-E2E Online expansion tests for MiniCPM-o 4.5 covering modality combinations,
-in-process token2wav vocoder behavior, max_num_seqs=1 serial execution,
-and edge-case validation.
+"""E2E online expansion tests for MiniCPM-o 4.5.
+
+These cover modality combinations, separate Code2Wav behavior, sequential
+request isolation, longer audio output, and Chinese speech.
+
+Like Qwen3-Omni expansion, ``default`` exercises ``--no-async-chunk`` and
+``async_chunk`` exercises ``--async-chunk``.
 """
 
 import os
@@ -18,20 +21,29 @@ pytestmark = [pytest.mark.full_model, pytest.mark.omni]
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
 _MODEL = "openbmb/MiniCPM-o-4_5"
-_CI_DEPLOY = get_deploy_config_path("minicpmo_4_5_batching.yaml")
+_CI_DEPLOY = get_deploy_config_path("minicpmo_4_5.yaml")
 
+# Deploy YAML defaults to ``async_chunk: true``; keep sync and async as
+# separate parametrizations so both handoff paths are exercised.
 test_params = [
     pytest.param(
         OmniServerParams(
             model=_MODEL,
             stage_config_path=_CI_DEPLOY,
-            use_stage_cli=True,
-            server_args=[
-                "--trust-remote-code",
-            ],
+            use_stage_cli=False,
+            server_args=["--trust-remote-code", "--no-async-chunk"],
         ),
         id="default",
-    )
+    ),
+    pytest.param(
+        OmniServerParams(
+            model=_MODEL,
+            stage_config_path=_CI_DEPLOY,
+            use_stage_cli=False,
+            server_args=["--trust-remote-code", "--async-chunk"],
+        ),
+        id="async_chunk",
+    ),
 ]
 
 
@@ -63,12 +75,12 @@ def get_max_batch_size(size_type="few"):
     return batch_sizes.get(size_type, 5)
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_text_video_to_text_001(omni_server, openai_client) -> None:
     """
     Test text + video input generating text output.
-    Deploy Setting: default 2GPU
+    Deploy Setting: default single GPU
     Input Modal: text + video
     Output Modal: text
     Input Setting: stream=False
@@ -90,15 +102,13 @@ def test_text_video_to_text_001(omni_server, openai_client) -> None:
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_sequential_requests_independent(omni_server, openai_client) -> None:
     """
-    Verify that sequential requests produce independent results and state
-    does not leak between requests. MiniCPM-o 4.5 has max_num_seqs=1 so
-    requests are processed one-at-a-time; the second request must not
-    receive the first request's audio or token state.
-    Deploy Setting: default 2GPU
+    Verify that sequential requests produce independent results and that the
+    second request does not receive the first request's audio or token state.
+    Deploy Setting: default single GPU
     Input Modal: text (two different prompts)
     Output Modal: text + audio (both)
     """
@@ -133,33 +143,42 @@ def test_sequential_requests_independent(omni_server, openai_client) -> None:
     )
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_text_to_audio_long_output_001(omni_server, openai_client) -> None:
     """
     Test text input generating a longer audio output to exercise the
-    token2wav decoder across multiple frames.
-    Deploy Setting: default 2GPU
+    Code2Wav stage across multiple frames.
+    Deploy Setting: default single GPU
     Input Modal: text (longer prompt)
     Output Modal: text + audio
     Input Setting: stream=True
     """
     messages = dummy_messages_from_mix_data(
         system_prompt=get_system_prompt(),
-        content_text="Tell me a short story about a cat in about 50 words.",
+        content_text="What is the capital of China? Answer in 40 words.",
     )
 
     request_config = {
         "model": omni_server.model,
         "messages": messages,
+        "modalities": ["text", "audio"],
         "stream": True,
-        "key_words": {"audio": ["cat"]},
+        # Delimit the assistant answer with the TTS template so the Talker
+        # does not spend its codec-token budget speaking hidden reasoning.
+        "extra_body": {
+            "chat_template_kwargs": {
+                "use_tts_template": True,
+                "enable_thinking": False,
+            }
+        },
+        "key_words": {"audio": ["Beijing"]},
     }
 
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
+@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_chinese_text_to_audio(omni_server, openai_client) -> None:
     """

tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py

@@ -12,14 +12,11 @@ import pytest
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
-DEMO_PATH = Path(__file__).resolve().parents[1] / "e2e/online_serving/minicpmo_realtime_duplex_scenarios.py"
-MULTI_DEMO_PATH = (
-    Path(__file__).resolve().parents[1] / "e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py"
-)
-PAIR_DEMO_PATH = Path(__file__).resolve().parents[1] / "e2e/online_serving/run_minicpmo_realtime_duplex_demo_pair.py"
-SOFT_INTERRUPT_DEMO_PATH = (
-    Path(__file__).resolve().parents[1] / "e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py"
-)
+DRIVER_DIR = Path(__file__).resolve().parent
+DEMO_PATH = DRIVER_DIR / "minicpmo_realtime_duplex_scenarios.py"
+MULTI_DEMO_PATH = DRIVER_DIR / "run_minicpmo_realtime_duplex_multi_session.py"
+PAIR_DEMO_PATH = DRIVER_DIR / "run_minicpmo_realtime_duplex_demo_pair.py"
+SOFT_INTERRUPT_DEMO_PATH = DRIVER_DIR / "run_minicpmo_realtime_duplex_soft_interrupt.py"
 
 
 def _load_demo_module():
@@ -338,14 +335,7 @@ def test_realtime_duplex_soft_interrupt_accepts_explicit_ref_audio(monkeypatch):
     assert args.validation_mode == "model-policy"
 
 
-@pytest.mark.parametrize(
-    "extra_args",
-    [
-        ["--expect-second-response-substring", "一加一等于二"],
-        ["--input-sha256", "a" * 64],
-    ],
-)
-def test_realtime_duplex_soft_interrupt_response_required_needs_bound_fixture(monkeypatch, extra_args):
+def test_realtime_duplex_soft_interrupt_response_required_needs_bound_fixture(monkeypatch):
     demo = _load_soft_interrupt_demo_module()
     monkeypatch.setattr(
         demo.sys,
@@ -360,7 +350,6 @@ def test_realtime_duplex_soft_interrupt_response_required_needs_bound_fixture(mo
             "ref.wav",
             "--validation-mode",
             "response-required",
-            *extra_args,
         ],
     )
 
@@ -412,7 +401,6 @@ def test_realtime_duplex_soft_interrupt_accepts_multi_delta_handoff_sequence(tmp
             "_client_received_at_s": 2.6,
         },
         {"type": "response.done", "response_id": first_response_id, "_client_received_at_s": 3.0},
-        {"type": "response.listen", "_client_received_at_s": 3.2},
         {"type": "response.created", "response": {"id": second_response_id}, "_client_received_at_s": 4.0},
         {"type": "response.speak", "response_id": second_response_id, "_client_received_at_s": 4.0},
         {
@@ -458,10 +446,11 @@ def test_realtime_duplex_soft_interrupt_accepts_multi_delta_handoff_sequence(tmp
         validation_mode="response-required",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring="一加一等于二",
+        expect_followup_response_substring="一加一等于二",
     )
 
     assert summary["ok"] is True
+    assert summary["listen_between_responses"] is False
     assert summary["second_response_before_final_commit"] is True
     assert summary["final_listen_after_commit"] is True
     assert summary["response_audio_contract_ok"] is True
@@ -498,7 +487,7 @@ def test_realtime_duplex_soft_interrupt_model_policy_accepts_single_response(tmp
         validation_mode="model-policy",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring=None,
+        expect_followup_response_substring=None,
     )
 
     assert summary["ok"] is True
@@ -537,14 +526,14 @@ def test_realtime_duplex_soft_interrupt_response_required_rejects_single_respons
         validation_mode="response-required",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring="一加一等于二",
+        expect_followup_response_substring="一加一等于二",
     )
 
     assert summary["ok"] is False
     assert summary["enough_responses"] is False
 
 
-def test_realtime_duplex_soft_interrupt_checks_expected_text_in_second_response(tmp_path):
+def test_realtime_duplex_soft_interrupt_reports_text_expectation_without_gating(tmp_path):
     demo = _load_soft_interrupt_demo_module()
     output = tmp_path / "second_response_text"
     output.mkdir()
@@ -611,11 +600,32 @@ def test_realtime_duplex_soft_interrupt_checks_expected_text_in_second_response(
         validation_mode="response-required",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring="一加一等于二",
+        expect_followup_response_substring="一加一等于二",
     )
 
-    assert summary["ok"] is False
-    assert summary["second_response_transcript_expectation_ok"] is False
+    assert summary["ok"] is True
+    assert summary["followup_response_transcript_ok"] is True
+    assert summary["followup_response_transcript_expectation_ok"] is False
+
+    events = [
+        event
+        for event in events
+        if event.get("response_id") != second_response_id or event["type"] != "response.audio_transcript.delta"
+    ]
+    (output / "events.jsonl").write_text(
+        "".join(demo.json.dumps(event) + "\n" for event in events),
+        encoding="utf-8",
+    )
+    missing_transcript_summary = demo.summarize_artifacts(
+        output_dir=output,
+        validation_mode="response-required",
+        min_responses=2,
+        min_audio_deltas_per_response=2,
+        expect_followup_response_substring="一加一等于二",
+    )
+
+    assert missing_transcript_summary["ok"] is False
+    assert missing_transcript_summary["followup_response_transcript_ok"] is False
 
 
 def test_realtime_duplex_multi_session_resume_url_disables_autostart():
@@ -789,6 +799,7 @@ def test_realtime_duplex_demo_reads_response_playback_cursor():
 def test_realtime_duplex_demo_partitions_timing_by_response_identity():
     demo = _load_demo_module()
     state = demo.DemoState()
+    state.input_commit_sent_at_s.extend((9.9, 19.9))
     for response_id, created_at_s, audio_at_s, token_count in (
         ("resp-1", 10.0, 10.1, 3),
         ("resp-2", 20.0, 20.2, 5),
@@ -816,15 +827,66 @@ def test_realtime_duplex_demo_partitions_timing_by_response_identity():
             },
             received_at_s=audio_at_s,
         )
+        state.add(
+            {
+                "type": "response.audio_transcript.delta",
+                "response_id": response_id,
+                "delta": "audio",
+            },
+            received_at_s=created_at_s + 0.05,
+        )
 
     timings = state.response_timing_summaries()
+    requests = state.session_request_metrics(session_id="seed-tts-session")
+    session_metrics = state.session_metric_summary(session_id="seed-tts-session")
 
     assert timings["resp-1"]["stage0_tokens"]["output_token_count"] == 3
     assert timings["resp-1"]["audio_output"]["response_created_to_first_audio_ms"] == 100.0
-    assert timings["resp-1"]["audio_output"]["commit_to_first_audio_ms"] is None
+    assert timings["resp-1"]["audio_output"]["commit_to_first_audio_ms"] == 200.0
     assert timings["resp-2"]["stage0_tokens"]["output_token_count"] == 5
     assert timings["resp-2"]["audio_output"]["response_created_to_first_audio_ms"] == 200.0
-    assert timings["resp-2"]["audio_output"]["commit_to_first_audio_ms"] is None
+    assert timings["resp-2"]["audio_output"]["commit_to_first_audio_ms"] == 300.0
+    assert requests == [
+        {
+            "session_id": "seed-tts-session",
+            "request_index": 0,
+            "response_id": "resp-1",
+            "ttft_ms": 150.0,
+            "ttfp_ms": 200.0,
+            "rtf": 2.5,
+            "audio_generation_ms": 200.0,
+            "audio_duration_ms": 80.0,
+            "source": "client_monotonic_receive",
+            "measurement_origin": {
+                "ttft": "input_audio_buffer.commit client send to first non-empty text delta",
+                "ttfp": "input_audio_buffer.commit client send to first audio packet",
+                "rtf": "commit-to-last-audio receive time divided by emitted audio duration",
+            },
+        },
+        {
+            "session_id": "seed-tts-session",
+            "request_index": 1,
+            "response_id": "resp-2",
+            "ttft_ms": 150.0,
+            "ttfp_ms": 300.0,
+            "rtf": 3.75,
+            "audio_generation_ms": 300.0,
+            "audio_duration_ms": 80.0,
+            "source": "client_monotonic_receive",
+            "measurement_origin": {
+                "ttft": "input_audio_buffer.commit client send to first non-empty text delta",
+                "ttfp": "input_audio_buffer.commit client send to first audio packet",
+                "rtf": "commit-to-last-audio receive time divided by emitted audio duration",
+            },
+        },
+    ]
+    assert session_metrics == {
+        "session_id": "seed-tts-session",
+        "audio_turn_count": 2,
+        "mean_ttft_ms": 150.0,
+        "mean_ttfp_ms": 250.0,
+        "mean_rtf": 3.125,
+    }
 
 
 def test_realtime_duplex_demo_model_policy_accepts_one_listen_per_streamed_turn():
@@ -1523,7 +1585,7 @@ def test_realtime_duplex_demo_waits_at_each_model_unit_and_stops_after_speak():
     assert len(ws.messages) == 10
 
 
-def test_realtime_duplex_demo_listen_only_overlap_sends_next_turn_before_first_done(monkeypatch):
+def test_realtime_duplex_demo_listen_only_overlap_accepts_silence_unit_before_first_done(monkeypatch):
     demo = _load_demo_module()
     state = demo.DemoState()
     send_calls = 0
@@ -1548,7 +1610,6 @@ def test_realtime_duplex_demo_listen_only_overlap_sends_next_turn_before_first_d
         del args
         assert not state.response_done("resp-first")
         kwargs["on_model_unit_ready"]()
-        state.add({"type": "input_audio_buffer.speech_started", "turn": 2})
         _add_response_transcript(state, "resp-first", transcript="第一轮完成", audio=False)
         state.add({"type": "response.created", "response": {"id": "resp-second"}})
         _add_response_transcript(state, "resp-second", transcript="第二轮完成", audio=True)
@@ -1568,7 +1629,7 @@ def test_realtime_duplex_demo_listen_only_overlap_sends_next_turn_before_first_d
             ws,
             state,
             b"\x00\x00",
-            b"\x01\x00",
+            b"\x00\x00",
             transcripts=("first", "second"),
             durations_ms=(None, None),
             chunk_ms=200,

tests/engine/test_arg_utils.py

@@ -300,6 +300,56 @@ def test_strip_single_engine_args_model_does_not_trigger_warning(mocker):
     assert "model" not in warned_args
 
 
+def test_new_format_deploy_yaml_does_not_strip_interleave_cli_overrides(tmp_path):
+    """New-format deploy YAMLs must not strip interleave_mm_strings / media_io_kwargs.
+
+    Pytest historically passed ``--stage-configs-path`` for deploy YAMLs; stripping
+    those CLI flags dropped MiniCPM interleaved AV packing and cut Daily-Omni
+    accuracy from ~78% to ~70%.
+    """
+    from vllm_omni.entrypoints.utils import is_new_format_deploy_config
+
+    yaml_path = tmp_path / "minicpm_deploy.yaml"
+    yaml_path.write_text(
+        """
+pipeline: minicpmo_4_5
+async_chunk: true
+stages:
+  - stage_id: 0
+    devices: "0"
+""".strip(),
+        encoding="utf-8",
+    )
+    path = str(yaml_path)
+    assert is_new_format_deploy_config(path)
+
+    kwargs = {
+        "stage_configs_path": path,
+        "trust_remote_code": True,
+        "interleave_mm_strings": True,
+        "media_io_kwargs": {"video": {"fps": 1, "num_frames": 128}},
+        "gpu_memory_utilization": 0.55,
+    }
+    # Mirror AsyncOmniEngine._resolve_stage_configs strip decision.
+    if kwargs.get("stage_configs_path") is not None and not is_new_format_deploy_config(kwargs["stage_configs_path"]):
+        base = AsyncOmniEngine._strip_single_engine_args(dict(kwargs))
+    else:
+        base = dict(kwargs)
+    assert base.get("interleave_mm_strings") is True
+    assert base.get("media_io_kwargs") == {"video": {"fps": 1, "num_frames": 128}}
+    # Parent EngineArgs fields also remain available for deploy merge.
+    assert base.get("gpu_memory_utilization") == 0.55
+
+    legacy = tmp_path / "legacy.yaml"
+    legacy.write_text("stage_args:\n  - stage_id: 0\n", encoding="utf-8")
+    assert not is_new_format_deploy_config(str(legacy))
+    legacy_kwargs = dict(kwargs, stage_configs_path=str(legacy))
+    stripped = AsyncOmniEngine._strip_single_engine_args(legacy_kwargs)
+    assert "interleave_mm_strings" not in stripped
+    assert "media_io_kwargs" not in stripped
+    assert "gpu_memory_utilization" not in stripped
+
+
 # For https://github.com/vllm-project/vllm-omni/issues/3293
 def test_tensor_parallel_size_none_is_handled():
     """Ensure the tensor parallel size of None isn't forwarded."""

tests/engine/test_multimodal_accumulation.py

@@ -18,6 +18,7 @@ def test_chunk_accumulation_policy_replaces_snapshots_and_drains_delta_state():
     accumulated = MultimodalPayload.from_dict(
         {
             "audio": torch.tensor([1.0]),
+            "meta.segment_end": torch.tensor([0]),
             "meta.tts_is_last_chunk": torch.tensor([0]),
             "meta.turn_end": torch.tensor([0]),
             "meta.stable_request_value": "keep",
@@ -26,6 +27,7 @@ def test_chunk_accumulation_policy_replaces_snapshots_and_drains_delta_state():
     incoming = MultimodalPayload.from_dict(
         {
             "audio": torch.tensor([2.0]),
+            "meta.segment_end": torch.tensor([1]),
             "meta.tts_is_last_chunk": torch.tensor([1]),
             "meta.turn_end": torch.tensor([1]),
         }
@@ -41,6 +43,7 @@ def test_chunk_accumulation_policy_replaces_snapshots_and_drains_delta_state():
     drain_delta_payload(merged)
 
     assert "audio" not in merged
+    assert "meta.segment_end" not in merged
     assert "meta.tts_is_last_chunk" not in merged
     assert "meta.turn_end" not in merged
     assert merged.metadata["meta.stable_request_value"] == "keep"

tests/engine/test_orchestrator.py

@@ -1813,6 +1813,56 @@ async def test_stage_pool_submit_update_refreshes_output_processor_state() -> No
 
 
 @pytest.mark.asyncio
+async def test_handle_streaming_update_unknown_request_is_dropped() -> None:
+    """A streaming update for an untracked request must not resurrect it.
+
+    Updates always follow the first-chunk add_request through the same
+    ordered submission queue, so an unknown id can only mean the request
+    already finished or was aborted (e.g. the client disconnected).
+    Falling back to add_request created headless sessions that kept
+    cycling through the stages (issue #4271).
+    """
+
+    class RecordingPool:
+        def __init__(self) -> None:
+            self.calls: list[tuple[str, Any]] = []
+
+        async def submit_update(self, request_id, req_state, request, *, prompt_text=None) -> int:
+            self.calls.append((request_id, prompt_text))
+            return 0
+
+    pool = RecordingPool()
+    orchestrator = object.__new__(Orchestrator)
+    orchestrator.async_chunk = False
+    orchestrator.request_states = {}
+    orchestrator.stage_pools = [pool]
+    add_request_calls: list[str] = []
+
+    async def record_add_request(msg) -> None:
+        add_request_calls.append(msg.request_id)
+
+    orchestrator._handle_add_request = record_add_request
+
+    await orchestrator._handle_streaming_update(
+        StageSubmissionMessage(
+            type="streaming_update",
+            request_id="req-unknown",
+            prompt=SimpleNamespace(request_id="req-unknown", prompt_token_ids=[1], resumable=True),
+            original_prompt={"prompt": "segment-2"},
+            output_prompt_text="segment-2",
+            sampling_params_list=[_sampling_params()],
+            final_stage_id=0,
+            preprocess_ms=0.0,
+            request_timestamp=time.time(),
+            enqueue_ts=time.perf_counter(),
+        )
+    )
+
+    assert pool.calls == []
+    assert add_request_calls == []
+    assert "req-unknown" not in orchestrator.request_states
+
+
 async def test_handle_streaming_update_passes_prompt_text_to_stage_pool() -> None:
     class RecordingPool:
         def __init__(self) -> None:

tests/engine/test_orchestrator_stage_input_bridge.py

@@ -7,15 +7,24 @@ import asyncio
 import queue
 from types import SimpleNamespace
 from typing import Any
-from unittest.mock import AsyncMock
+from unittest.mock import ANY, AsyncMock, MagicMock
 
 import janus
 import pytest
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import SamplingParams
 
-from vllm_omni.engine.orchestrator import Orchestrator, OrchestratorRequestState
+from vllm_omni.engine.orchestrator import (
+    Orchestrator,
+    OrchestratorRequestState,
+    _OrchestratorDuplexStagePort,
+)
 from vllm_omni.engine.stage_pool import StagePool
+from vllm_omni.experimental.fullduplex.engine.contracts import (
+    DuplexStageRequestContext,
+    DuplexStageSubmission,
+)
+from vllm_omni.experimental.fullduplex.engine.messages import DuplexFence
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -107,11 +116,50 @@ class FakePrewarmPool:
 
     async def submit_initial(self, _request_id, _req_state, request, prompt_text=None):
         self.submitted.append(request)
+        return 0
 
     def get_bound_replica_id(self, _request_id):
         return 0
 
 
+def _duplex_stage_port_submission():
+    stage_pools = []
+    for stage_id in range(3):
+        pool = SimpleNamespace(
+            stage_client=SimpleNamespace(default_sampling_params=SamplingParams(max_tokens=1)),
+            stage_vllm_config=SimpleNamespace(model_config=SimpleNamespace(max_model_len=64)),
+            submit_initial=AsyncMock(return_value=stage_id + 10),
+            submit_update=AsyncMock(return_value=stage_id + 20),
+        )
+        stage_pools.append(pool)
+    request_states: dict[str, OrchestratorRequestState] = {}
+    prewarm = AsyncMock()
+    port = _OrchestratorDuplexStagePort(
+        stage_pools=stage_pools,
+        request_states=request_states,
+        running_counter=None,
+        cleanup_request_ids=AsyncMock(),
+        async_chunk=True,
+        prewarm_async_chunk_stages=prewarm,
+    )
+    context = DuplexStageRequestContext(
+        request_id="req-duplex",
+        session_id="session-duplex",
+        fence=DuplexFence("session-duplex"),
+        stage_id=0,
+        final_stage_id=2,
+        config_generation=0,
+        sampling_params=tuple(SamplingParams(max_tokens=1) for _ in range(3)),
+    )
+    port.ensure_request(context)
+    submission = DuplexStageSubmission(
+        context=context,
+        prompt={"prompt_token_ids": [1, 2]},
+        already_submitted=False,
+    )
+    return port, stage_pools, request_states, prewarm, submission
+
+
 def _request_output(request_id: str) -> RequestOutput:
     completion = CompletionOutput(
         index=0,
@@ -199,11 +247,13 @@ async def test_async_prewarm_skips_outgoing_only_stage() -> None:
     stage2 = FakePrewarmPool("receiver")
     orchestrator.stage_pools = [stage0, stage1, stage2]
     orchestrator._emit_tx_edge = lambda **_kwargs: None
+    orchestrator._record_duplex_stage_submission = MagicMock()
     req_state = OrchestratorRequestState(
         request_id="req-prewarm",
         prompt={"prompt_token_ids": [1, 2]},
         sampling_params_list=[SamplingParams(max_tokens=1) for _ in range(3)],
         final_stage_id=2,
+        duplex_identity=SimpleNamespace(),
     )
 
     await orchestrator._prewarm_async_chunk_stages(
@@ -216,6 +266,23 @@ async def test_async_prewarm_skips_outgoing_only_stage() -> None:
     assert len(stage2.submitted) == 1
     assert 1 not in req_state.stage_submit_ts
     assert 2 in req_state.stage_submit_ts
+    orchestrator._record_duplex_stage_submission.assert_called_once_with(
+        2,
+        "req-prewarm",
+        0,
+        req_state,
+    )
+
+
+@pytest.mark.asyncio
+async def test_duplex_prewarm_runs_after_first_stage0_submission() -> None:
+    port, stage_pools, request_states, prewarm, submission = _duplex_stage_port_submission()
+
+    result = await port.submit(submission)
+
+    assert result.stage_id == 0
+    stage_pools[0].submit_initial.assert_awaited_once()
+    prewarm.assert_awaited_once_with("req-duplex", ANY, request_states["req-duplex"])
 
 
 @pytest.mark.asyncio

tests/engine/test_output_processor.py

@@ -14,7 +14,7 @@ from vllm.v1.engine import FinishReason
 from vllm.v1.engine.output_processor import OutputProcessor as VLLMOutputProcessor
 
 from vllm_omni.outputs import output_processor
-from vllm_omni.outputs.output_modality import OutputModalityNames
+from vllm_omni.outputs.output_modality import OutputModality, OutputModalityNames
 from vllm_omni.outputs.output_processor import MultimodalOutputProcessor, OmniRequestState
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -452,6 +452,127 @@ def test_no_detokenizer_make_request_output():
     assert AUDIO in result.outputs[0].multimodal_output
 
 
+def test_no_detokenizer_process_outputs_returns_nonterminal_audio_chunk(monkeypatch):
+    """Generation-stage MM-only chunks must reach the orchestrator.
+
+    StagePool registers generation requests without a per-request collector,
+    so the processor return value is the only path from Stage2 to
+    Orchestrator._handle_processed_outputs().
+    """
+    state = _make_no_detok_state(RequestOutputKind.DELTA)
+    processor = object.__new__(MultimodalOutputProcessor)
+    processor.output_modality = OutputModality.AUDIO
+    processor.request_states = {"r": state}
+    upstream_result = SimpleNamespace(request_outputs=[], reqs_to_abort=[])
+    monkeypatch.setattr(
+        VLLMOutputProcessor,
+        "process_outputs",
+        lambda *_args, **_kwargs: upstream_result,
+    )
+    engine_output = SimpleNamespace(
+        request_id="r",
+        multimodal_output={
+            "model_outputs": torch.arange(8, dtype=torch.float32),
+            "sr": torch.tensor(24000, dtype=torch.int32),
+        },
+        output_type="audio",
+        pooling_output=None,
+        new_token_ids=[],
+        finish_reason=None,
+        stop_reason=None,
+        kv_transfer_params=None,
+        routed_experts=None,
+        num_cached_tokens=0,
+    )
+
+    result = processor.process_outputs([engine_output])
+
+    assert len(result.request_outputs) == 1
+    assert result.request_outputs[0].request_id == "r"
+    assert AUDIO in result.request_outputs[0].outputs[0].multimodal_output
+
+
+def _make_mm_only_output_processor(monkeypatch):
+    processor = object.__new__(MultimodalOutputProcessor)
+    processor.output_modality = OutputModality.AUDIO
+    processor.request_states = {"r": _make_no_detok_state(RequestOutputKind.DELTA)}
+    monkeypatch.setattr(
+        VLLMOutputProcessor,
+        "process_outputs",
+        lambda *_args, **_kwargs: SimpleNamespace(
+            request_outputs=[],
+            reqs_to_abort=[],
+        ),
+    )
+
+    def finish_request(req_state):
+        processor.request_states.pop(req_state.request_id, None)
+
+    processor._finish_request = finish_request
+    return processor
+
+
+def _audio_engine_output(*, is_segment_finished: bool, is_last_chunk: bool):
+    return SimpleNamespace(
+        request_id="r",
+        multimodal_output={
+            "model_outputs": torch.arange(8, dtype=torch.float32),
+            "meta.tts_is_last_chunk": torch.tensor([int(is_last_chunk)]),
+        },
+        output_type="audio",
+        pooling_output=None,
+        new_token_ids=[],
+        finish_reason=FinishReason.STOP,
+        stop_reason=None,
+        kv_transfer_params=None,
+        routed_experts=None,
+        num_cached_tokens=0,
+        is_segment_finished=is_segment_finished,
+    )
+
+
+def test_mm_only_segment_finish_retains_request_state_for_next_audio_chunk(monkeypatch):
+    processor = _make_mm_only_output_processor(monkeypatch)
+
+    first = processor.process_outputs([_audio_engine_output(is_segment_finished=True, is_last_chunk=False)])
+
+    assert "r" in processor.request_states
+    assert len(first.request_outputs) == 1
+    assert first.request_outputs[0].finished is False
+
+    second = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=True)])
+
+    assert len(second.request_outputs) == 1
+    assert second.request_outputs[0].finished is True
+    assert "r" not in processor.request_states
+
+
+def test_mm_only_nonfinal_audio_retains_state_without_segment_marker(monkeypatch):
+    processor = _make_mm_only_output_processor(monkeypatch)
+
+    first = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=False)])
+
+    assert "r" in processor.request_states
+    assert len(first.request_outputs) == 1
+    assert first.request_outputs[0].finished is False
+
+    second = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=True)])
+
+    assert len(second.request_outputs) == 1
+    assert second.request_outputs[0].finished is True
+    assert "r" not in processor.request_states
+
+
+def test_mm_only_terminal_finish_removes_request_state(monkeypatch):
+    processor = _make_mm_only_output_processor(monkeypatch)
+
+    result = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=True)])
+
+    assert len(result.request_outputs) == 1
+    assert result.request_outputs[0].finished is True
+    assert "r" not in processor.request_states
+
+
 def test_no_detokenizer_make_request_output_with_routed_experts():
     """make_request_output accepts the routed_experts arg that the multimodal
     output channel (_process_mm_only_outputs) passes positionally, and attaches

tests/engine/test_single_stage_mode.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import os
+import socket
 from contextlib import contextmanager
 from types import SimpleNamespace
 from typing import Any
@@ -11,6 +12,7 @@ import pytest
 from pytest_mock import MockerFixture
 from vllm.v1.engine.utils import EngineZmqAddresses
 
+from vllm_omni.config.stage_config import DuplexSessionRuntimeConfig
 from vllm_omni.engine.async_omni_engine import AsyncOmniEngine
 from vllm_omni.engine.stage_engine_core_client import StageEngineCoreClientBase
 from vllm_omni.engine.stage_engine_startup import (
@@ -189,6 +191,34 @@ class TestOmniMasterServerAllocation:
         output_port = int(alloc.output_bind_address.split(":")[-1])
         assert len({handshake_port, input_port, output_port}) == 3
 
+    def test_route_ports_remain_os_reserved_until_master_releases_them(self):
+        server = OmniMasterServer(master_address="127.0.0.1", master_port=15005, stage_ids=[0])
+        alloc = server.get_allocation(0)
+        ports = [
+            int(alloc.handshake_bind_address.rsplit(":", 1)[-1]),
+            int(alloc.input_bind_address.rsplit(":", 1)[-1]),
+            int(alloc.output_bind_address.rsplit(":", 1)[-1]),
+        ]
+
+        for port in ports:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                with pytest.raises(OSError):
+                    probe.bind(("127.0.0.1", port))
+
+        server.release_route_port_reservations(0, handshake=True, data=False)
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+            probe.bind(("127.0.0.1", ports[0]))
+        for port in ports[1:]:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                with pytest.raises(OSError):
+                    probe.bind(("127.0.0.1", port))
+
+        server.stop()
+
+        for port in ports[1:]:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                probe.bind(("127.0.0.1", port))
+
     def test_get_zmq_addresses_returns_bind_addresses(self):
         server = OmniMasterServer(master_address="127.0.0.1", master_port=15002, stage_ids=[0])
         alloc = server.get_allocation(0)
@@ -397,6 +427,25 @@ class TestSingleStageModeDetection:
         assert engine.single_stage_mode is False
         assert engine._single_stage_id_filter is None
 
+    def test_stage_configs_path_loads_duplex_runtime_config(self, mocker: MockerFixture):
+        duplex_session = DuplexSessionRuntimeConfig(max_sessions=2)
+        mocker.patch(
+            "vllm_omni.engine.async_omni_engine.StageConfigFactory.get_pipeline_config",
+            return_value=None,
+        )
+        load_deploy_config = mocker.patch(
+            "vllm_omni.engine.async_omni_engine.load_deploy_config",
+            return_value=SimpleNamespace(duplex_session=duplex_session),
+        )
+
+        engine = self._make_engine_no_thread(
+            mocker,
+            stage_configs_path="/fake/duplex.yaml",
+        )
+
+        load_deploy_config.assert_called_once_with("/fake/duplex.yaml")
+        assert engine.duplex_session_config is duplex_session
+
     def test_single_stage_mode_without_stage_id_has_no_filter(self, mocker: MockerFixture):
         engine = self._make_engine_no_thread(
             mocker,

tests/entrypoints/openai_api/test_duplex_handler.py

@@ -51,6 +51,19 @@ from vllm_omni.outputs import OmniRequestOutput
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
 
+def test_native_input_append_supports_explicit_session_opt_out():
+    session = DuplexSession(
+        "sid-explicit-tts",
+        DuplexSessionConfig(model="test-model"),
+        capabilities=DuplexCapabilities.minicpmo45_native(),
+    )
+
+    assert OmniDuplexSessionHandler._uses_native_input_append(session) is True
+
+    session.config.extra_body["minicpmo45_native_duplex"] = False
+    assert OmniDuplexSessionHandler._uses_native_input_append(session) is False
+
+
 class _ModelConfig:
     model = "test-model"
 
@@ -2520,6 +2533,52 @@ def test_duplex_auto_response_tts_scheduler_eos_fallback():
     assert results[0]["abort_data_plane_request"] is True
 
 
+def test_duplex_auto_response_rearms_tts_boundary_for_next_segment_in_same_turn():
+    data_plane = _test_data_plane()
+    request_id = "duplex-sid-next-segment-same-turn-e0-stage0"
+    session = DuplexSession(
+        session_id="sid-next-segment-same-turn",
+        config=DuplexSessionConfig(extra_body={"auto_response": True}),
+    )
+    session.bind_response_turn(0)
+    data_plane.begin_request(request_id)
+
+    first_segment = _duplex_tts_output(
+        request_id=request_id,
+        samples=24000,
+        finished=False,
+        tts_is_last_chunk=True,
+        turn_id=0,
+    )
+    first_results = _project_data_plane(
+        data_plane,
+        {"data_plane_outputs": [first_segment]},
+        session=session,
+    )
+
+    assert len(first_results) == 1
+    assert first_results[0]["abort_data_plane_request"] is True
+
+    data_plane.begin_request(request_id)
+    second_segment = _duplex_tts_output(
+        request_id=request_id,
+        samples=48000,
+        finished=False,
+        text="hello again",
+        tts_is_last_chunk=True,
+        turn_id=0,
+    )
+    second_results = _project_data_plane(
+        data_plane,
+        {"data_plane_outputs": [second_segment]},
+        session=session,
+    )
+
+    assert len(second_results) == 1
+    assert second_results[0]["audio_data"] == "wav-24000"
+    assert second_results[0]["abort_data_plane_request"] is True
+
+
 def test_duplex_turn_end_is_not_swallowed_by_finished_segment_fallback():
     data_plane = _test_data_plane()
     request_id = "duplex-sid-turn-end-after-segment-e0-stage0"
@@ -6202,19 +6261,27 @@ async def test_minicpmo_native_auto_response_keeps_request_bound_for_segment_con
         config_timeout_s=0.1,
         idle_timeout_s=1,
     )
-    ws = TimedWebSocket()
+
+    def on_send(ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") != "session.created":
+            return
+        session = handler._registry.get("sid-native-auto-continuation")
+        assert session is not None
+        session.capabilities.chunk_period_ms = 50
+        ws.put(
+            {
+                "type": "input_audio_buffer.append",
+                "audio": _pcm_f32_b64(800),
+                "format": "pcm_f32le",
+                "sample_rate_hz": 16000,
+            }
+        )
+        ws.put({"type": "input_audio_buffer.commit", "final": True})
+
+    ws = TimedWebSocket(on_send=on_send)
     event = _native_session_create("sid-native-auto-continuation")
     event["session"]["extra_body"]["auto_response"] = True
     ws.put(event)
-    ws.put(
-        {
-            "type": "input_audio_buffer.append",
-            "audio": _pcm_f32_b64(16000),
-            "format": "pcm_f32le",
-            "sample_rate_hz": 16000,
-        }
-    )
-    ws.put({"type": "input_audio_buffer.commit", "final": True})
 
     await handler.handle_session(ws)
 
@@ -6225,6 +6292,112 @@ async def test_minicpmo_native_auto_response_keeps_request_bound_for_segment_con
     assert final is False
 
 
+@pytest.mark.asyncio
+async def test_minicpmo_native_auto_response_real_input_preempts_unsent_silence():
+    request_id = "duplex-sid-native-real-before-silence-e0-stage0"
+
+    class SequencedAppendEngine(FakeEngineClient):
+        def __init__(self) -> None:
+            control_result = {
+                "operation": "append",
+                "session_id": "sid-native-real-before-silence",
+                "ok": True,
+                "unsupported_count": 0,
+                "error_count": 0,
+                "stage_results": [
+                    {
+                        "result": {
+                            "data_plane_append": True,
+                            "request_id": request_id,
+                            "response_stage_id": 1,
+                        }
+                    }
+                ],
+            }
+            terminal_segment = _duplex_tts_output(
+                request_id=request_id,
+                samples=0,
+                finished=True,
+                tts_is_last_chunk=True,
+                token_ids=[151645],
+            )
+            super().__init__(
+                control_result=control_result,
+                collect_outputs=[[terminal_segment], [], []],
+                collect_delay_s=0.01,
+            )
+            self.append_sequence: list[str] = []
+            self.first_real_started = asyncio.Event()
+            self.second_real_started = asyncio.Event()
+            self.real_count = 0
+
+        async def append_duplex_input_async(self, session_id: str, **kwargs):
+            kwargs.pop("expected_epoch", None)
+            result = await super().append_duplex_input_async(session_id, **kwargs)
+            payload = kwargs.get("payload")
+            is_silence = (
+                isinstance(payload, dict)
+                and payload.get("audio") == OmniDuplexSessionHandler._NATIVE_SILENCE_UNIT_PAYLOAD_AUDIO
+            )
+            if is_silence:
+                self.append_sequence.append("silence")
+            else:
+                self.real_count += 1
+                self.append_sequence.append(f"real-{self.real_count}")
+                if self.real_count == 1:
+                    self.first_real_started.set()
+                elif self.real_count == 2:
+                    self.second_real_started.set()
+            return result
+
+    session_created = asyncio.Event()
+    engine = SequencedAppendEngine()
+    handler = OmniDuplexSessionHandler(
+        chat_service=FakeChatService(engine),
+        config_timeout_s=0.1,
+        idle_timeout_s=1,
+    )
+
+    def on_send(ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") != "session.created":
+            return
+        session = handler._registry.get("sid-native-real-before-silence")
+        assert session is not None
+        session.capabilities.chunk_period_ms = 50
+        ws.put(
+            {
+                "type": "input_audio_buffer.append",
+                "audio": _pcm_f32_b64(800),
+                "format": "pcm_f32le",
+                "sample_rate_hz": 16000,
+            }
+        )
+        session_created.set()
+
+    ws = TimedWebSocket(on_send=on_send, receive_timeout_s=2.0)
+    event = _native_session_create("sid-native-real-before-silence")
+    event["session"]["extra_body"]["auto_response"] = True
+    ws.put(event)
+    handler_task = asyncio.create_task(handler.handle_session(ws))
+
+    await asyncio.wait_for(session_created.wait(), timeout=1)
+    await asyncio.wait_for(engine.first_real_started.wait(), timeout=1)
+    await asyncio.sleep(0.02)
+    ws.put(
+        {
+            "type": "input_audio_buffer.append",
+            "audio": _pcm_f32_b64(800, value=0.07),
+            "format": "pcm_f32le",
+            "sample_rate_hz": 16000,
+        }
+    )
+    await asyncio.wait_for(engine.second_real_started.wait(), timeout=1)
+    ws.put({"type": "session.close"})
+    await asyncio.wait_for(handler_task, timeout=2)
+
+    assert engine.append_sequence[:2] == ["real-1", "real-2"]
+
+
 @pytest.mark.asyncio
 async def test_minicpmo_native_auto_response_real_input_waits_for_submitted_silence_tail():
     request_id = "duplex-sid-native-silence-tail-e0-stage0"
@@ -6294,18 +6467,26 @@ async def test_minicpmo_native_auto_response_real_input_waits_for_submitted_sile
         config_timeout_s=0.1,
         idle_timeout_s=1,
     )
-    ws = TimedWebSocket(receive_timeout_s=2.0)
+
+    def on_send(ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") != "session.created":
+            return
+        session = handler._registry.get("sid-native-silence-tail")
+        assert session is not None
+        session.capabilities.chunk_period_ms = 50
+        ws.put(
+            {
+                "type": "input_audio_buffer.append",
+                "audio": _pcm_f32_b64(800, value=0.05),
+                "format": "pcm_f32le",
+                "sample_rate_hz": 16000,
+            }
+        )
+
+    ws = TimedWebSocket(on_send=on_send, receive_timeout_s=2.0)
     event = _native_session_create("sid-native-silence-tail")
     event["session"]["extra_body"]["auto_response"] = True
     ws.put(event)
-    ws.put(
-        {
-            "type": "input_audio_buffer.append",
-            "audio": _pcm_f32_b64(16000, value=0.05),
-            "format": "pcm_f32le",
-            "sample_rate_hz": 16000,
-        }
-    )
     handler_task = asyncio.create_task(handler.handle_session(ws))
 
     await asyncio.wait_for(engine.silence_append_started.wait(), timeout=1)
@@ -6330,6 +6511,135 @@ async def test_minicpmo_native_auto_response_real_input_waits_for_submitted_sile
     assert engine.append_sequence == ["real-1", "silence-started", "silence-returned", "real-2"]
 
 
+@pytest.mark.asyncio
+async def test_minicpmo_native_auto_response_preserves_silence_continuations_across_append_tails():
+    request_id = "duplex-sid-native-continuation-tails-e0-stage0"
+
+    class SequencedAppendEngine(FakeEngineClient):
+        def __init__(self) -> None:
+            control_result = {
+                "operation": "append",
+                "session_id": "sid-native-continuation-tails",
+                "ok": True,
+                "unsupported_count": 0,
+                "error_count": 0,
+                "stage_results": [
+                    {
+                        "result": {
+                            "data_plane_append": True,
+                            "request_id": request_id,
+                            "response_stage_id": 1,
+                        }
+                    }
+                ],
+            }
+            super().__init__(control_result=control_result)
+            self.real_append_started = asyncio.Event()
+            self.release_real_append = asyncio.Event()
+            self.first_silence_started = asyncio.Event()
+            self.release_first_silence = asyncio.Event()
+            self.second_silence_started = asyncio.Event()
+            self.release_collect = asyncio.Event()
+            self.silence_count = 0
+
+        async def append_duplex_input_async(self, session_id: str, **kwargs):
+            kwargs.pop("expected_epoch", None)
+            payload = kwargs.get("payload")
+            assert isinstance(payload, dict)
+            result = await super().append_duplex_input_async(session_id, **kwargs)
+            if payload.get("audio") != OmniDuplexSessionHandler._NATIVE_SILENCE_UNIT_PAYLOAD_AUDIO:
+                self.real_append_started.set()
+                await self.release_real_append.wait()
+                return result
+
+            self.silence_count += 1
+            if self.silence_count == 1:
+                self.first_silence_started.set()
+                await self.release_first_silence.wait()
+            else:
+                self.second_silence_started.set()
+            return result
+
+        async def collect_duplex_data_plane_outputs_async(
+            self,
+            request_id: str,
+            *,
+            response_stage_id: int | None = None,
+            timeout: float | None = None,
+        ) -> list[object]:
+            del timeout
+            self.collected.append((request_id, response_stage_id))
+            await self.release_collect.wait()
+            return []
+
+    session_created = asyncio.Event()
+
+    def on_send(_ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") == "session.created":
+            session_created.set()
+
+    engine = SequencedAppendEngine()
+    handler = OmniDuplexSessionHandler(
+        chat_service=FakeChatService(engine),
+        config_timeout_s=0.1,
+        idle_timeout_s=1,
+    )
+    ws = TimedWebSocket(on_send=on_send, receive_timeout_s=2.0)
+    event = _native_session_create("sid-native-continuation-tails")
+    event["session"]["extra_body"]["auto_response"] = True
+    ws.put(event)
+    ws.put(
+        {
+            "type": "input_audio_buffer.append",
+            "audio": _pcm_f32_b64(16000, value=0.05),
+            "format": "pcm_f32le",
+            "sample_rate_hz": 16000,
+        }
+    )
+    handler_task = asyncio.create_task(handler.handle_session(ws))
+
+    await asyncio.wait_for(session_created.wait(), timeout=1)
+    await asyncio.wait_for(engine.real_append_started.wait(), timeout=1)
+    session = handler._registry.get("sid-native-continuation-tails")
+    assert session is not None
+    session.capabilities.chunk_period_ms = 50
+    session.bind_request(request_id)
+    response_id = session.begin_response(turn_id=session.turn_id)
+    native = handler._minicpmo_session_state(session)
+    scheduler = native.silence_continuation_scheduler
+    assert scheduler is not None
+
+    payload = handler._native_silence_unit_payload()
+    payload["duplex_turn_id"] = session.turn_id
+    kwargs = {
+        "request_id": request_id,
+        "owner_id": f"response:{response_id}",
+        "response_id": response_id,
+        "response_owned": True,
+        "expected_epoch": session.epoch,
+        "expected_incarnation": session.incarnation,
+        "expected_model_turn_id": None,
+        "send_json": ws.send_json,
+    }
+    assert await scheduler(payload, **kwargs) is True
+
+    next_schedule = asyncio.create_task(scheduler(payload, **kwargs))
+    await asyncio.sleep(0.05)
+    assert not next_schedule.done()
+
+    engine.release_real_append.set()
+    await asyncio.wait_for(engine.first_silence_started.wait(), timeout=1)
+    assert not next_schedule.done()
+
+    engine.release_first_silence.set()
+    assert await asyncio.wait_for(next_schedule, timeout=1) is True
+    await asyncio.wait_for(engine.second_silence_started.wait(), timeout=1)
+    ws.put({"type": "session.close"})
+    await asyncio.wait_for(handler_task, timeout=2)
+
+    assert engine.silence_count == 2
+
+
 @pytest.mark.asyncio
 async def test_minicpmo_native_duplex_uses_segment_text_metadata_for_transcript_cursor():
     request_id = "duplex-sid-native-repeat-text-e0-stage0"

tests/entrypoints/test_serve.py

@@ -222,6 +222,7 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
     from vllm_omni.engine.stage_engine_startup import StageRegistrationResponse
 
     stage_cfg = _make_stage_cfg(0, stage_type="llm")
+    stage_cfg.engine_args["async_chunk"] = True
     parallel_config = SimpleNamespace(
         data_parallel_size_local=1,
         data_parallel_rank=0,
@@ -241,7 +242,10 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
         "vllm_omni.distributed.omni_connectors.utils.initialization.resolve_omni_kv_config_for_stage",
         return_value=(None, None, None),
     )
-    mocker.patch("vllm_omni.engine.stage_init_utils.get_stage_connector_spec", return_value={})
+    mock_connector_spec = mocker.patch(
+        "vllm_omni.engine.stage_init_utils.get_stage_connector_spec",
+        return_value={},
+    )
     mocker.patch("vllm_omni.engine.stage_init_utils.build_engine_args_dict", return_value={})
     mocker.patch(
         "vllm_omni.engine.stage_init_utils.build_vllm_config",
@@ -277,6 +281,7 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
     assert kwargs["omni_stage_config"] is stage_cfg
     assert kwargs["replica_id"] is None
     assert "socket_ownership" not in kwargs
+    assert mock_connector_spec.call_args.kwargs["async_chunk"] is True
 
     assert mock_manager_cls.call_count == 1
     mgr_kwargs = mock_manager_cls.call_args.kwargs

tests/entrypoints/test_utils.py

@@ -3,6 +3,8 @@
 import os
 from collections import Counter
 from dataclasses import dataclass
+from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 import torch
@@ -13,6 +15,10 @@ from vllm_omni.config.yaml_util import create_config
 from vllm_omni.diffusion.data import OmniDiffusionConfig
 from vllm_omni.engine.arg_utils import OmniEngineArgs
 from vllm_omni.engine.async_omni_engine import AsyncOmniEngine
+from vllm_omni.engine.stage_init_utils import (
+    get_stage_connector_spec,
+    load_omni_transfer_config_for_model,
+)
 from vllm_omni.entrypoints.utils import (
     _convert_dataclasses_to_dict,
     _filter_dict_like_object,
@@ -313,6 +319,58 @@ class TestResolveModelConfigPath:
         assert result is not None
         assert "glm_image.yaml" in result
 
+    def test_uses_registry_default_deploy_when_hf_model_type_differs(
+        self,
+        mocker: MockerFixture,
+        tmp_path,
+    ):
+        from vllm_omni.config.stage_config import _DEPLOY_DIR as real_deploy_dir
+        from vllm_omni.entrypoints import utils as utils_mod
+
+        hf_config = SimpleNamespace(
+            model_type="minicpmo",
+            architectures=["MiniCPMO"],
+            version="4.5",
+        )
+        model = "local/minicpmo-registry-default-test"
+        mocker.patch(
+            "vllm_omni.entrypoints.utils.get_config",
+            return_value=hf_config,
+        )
+        mocker.patch(
+            "vllm_omni.config.config_factory.get_config",
+            return_value=hf_config,
+        )
+        # A bare deploy/<hf_model_type>.yaml must not win over the registered
+        # pipeline default when HF model_type is not an OMNI_PIPELINES key.
+        (tmp_path / "minicpmo.yaml").write_text("stages: []\n", encoding="utf-8")
+        (tmp_path / "minicpmo_4_5.yaml").write_bytes((real_deploy_dir / "minicpmo_4_5.yaml").read_bytes())
+        mocker.patch.object(utils_mod, "_DEPLOY_DIR", tmp_path)
+
+        result = resolve_model_config_path(model)
+
+        assert result is not None
+        assert Path(result).as_posix().endswith("minicpmo_4_5.yaml")
+        assert Path(result).name == "minicpmo_4_5.yaml"
+        transfer_config = load_omni_transfer_config_for_model(model, result)
+        assert transfer_config is not None
+
+        sender = get_stage_connector_spec(
+            omni_transfer_config=transfer_config,
+            stage_id=1,
+            async_chunk=True,
+        )
+        receiver = get_stage_connector_spec(
+            omni_transfer_config=transfer_config,
+            stage_id=2,
+            async_chunk=True,
+        )
+        assert sender["extra"]["role"] == "sender"
+        assert receiver["extra"]["role"] == "receiver"
+        assert sender["extra"]["connector_get_sleep_s"] == 0.01
+        assert sender["extra"]["connector_get_max_wait_first_chunk"] == 3000
+        assert sender["extra"]["connector_get_max_wait"] == 300
+
 
 class TestLoadAndResolveStageConfigs:
     def test_load_and_resolve_with_kwargs(self, mocker: MockerFixture):

tests/examples/online_serving/test_minicpmo_streaming.py

@@ -1,29 +0,0 @@
-import base64
-import io
-import runpy
-import wave
-from pathlib import Path
-
-import pytest
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def test_append_wav_chunk_extracts_pcm_frames() -> None:
-    example = Path(__file__).parents[3] / "examples" / "online_serving" / "minicpmo" / "streaming_chat_completion.py"
-    append_wav_chunk = runpy.run_path(example)["append_wav_chunk"]
-    encoded_buffer = io.BytesIO()
-    with wave.open(encoded_buffer, "wb") as wav:
-        wav.setnchannels(1)
-        wav.setsampwidth(2)
-        wav.setframerate(24000)
-        wav.writeframes(b"\x01\x02\x03\x04")
-
-    pcm_parts: list[bytes] = []
-    audio_format = append_wav_chunk(
-        base64.b64encode(encoded_buffer.getvalue()).decode(),
-        pcm_parts,
-    )
-
-    assert audio_format == (1, 2, 24000)
-    assert pcm_parts == [b"\x01\x02\x03\x04"]

tests/examples/test_minicpmo_realtime_web_static.py

@@ -207,3 +207,68 @@ def test_playback_worklet_waits_before_playing_and_rebuffers_after_underrun():
         capture_output=True,
         text=True,
     )
+
+
+def test_playback_worklet_fades_terminal_drain_to_zero():
+    node = shutil.which("node")
+    if node is None:
+        pytest.skip("node is required for the AudioWorklet regression test")
+
+    script = textwrap.dedent(
+        """
+        const fs = require('fs');
+        const vm = require('vm');
+
+        global.sampleRate = 1000;
+        const messages = [];
+        global.AudioWorkletProcessor = class {
+          constructor() {
+            this.port = {
+              onmessage: null,
+              postMessage: (message) => messages.push(message),
+            };
+          }
+        };
+        let Processor = null;
+        global.registerProcessor = (_name, processor) => { Processor = processor; };
+        vm.runInThisContext(fs.readFileSync(process.argv[1], 'utf8'));
+
+        const processor = new Processor();
+        const pcm = new Int16Array(100);
+        pcm.fill(16384);
+        processor.handleMessage({
+          type: 'audio',
+          pcm,
+          responseId: 'response-terminal',
+          initialBufferMs: 400,
+        });
+        processor.handleMessage({
+          type: 'drain',
+          responseId: 'response-terminal',
+        });
+
+        const output = new Float32Array(100);
+        processor.process([], [[output]]);
+        const assert = (condition, message) => {
+          if (!condition) throw new Error(message);
+        };
+
+        assert(output.some((sample) => sample !== 0), 'terminal audio must still play');
+        assert(output[output.length - 1] === 0, 'terminal drain must fade the final sample to zero');
+        const tail = Array.from(output.slice(-5), Math.abs);
+        assert(
+          tail.every((sample, index) => index === 0 || sample <= tail[index - 1]),
+          'terminal drain fade must decrease monotonically',
+        );
+        assert(
+          messages.filter((message) => message.type === 'playback-drained').length === 1,
+          'terminal drain must be reported exactly once after playback',
+        );
+        """
+    )
+    subprocess.run(
+        [node, "-e", script, str(STATIC_ROOT / "playback_worklet.js")],
+        check=True,
+        capture_output=True,
+        text=True,
+    )

tests/fullduplex/minicpmo45/test_input.py

@@ -43,6 +43,54 @@ def test_append_emits_one_model_unit_when_multiple_units_are_buffered():
     assert buffer.pending_byte_count == 16_000 * 4
 
 
+def test_speech_marker_does_not_leak_across_irregular_chunk_boundaries():
+    buffer = MiniCPMO45PcmAppendBuffer()
+
+    assert buffer.append(pcm_payload(15_000), chunk_period_ms=1_000) is None
+    speech_unit = buffer.append(pcm_payload(2_000, speech=False), chunk_period_ms=1_000)
+    silence_unit = buffer.append(pcm_payload(15_000, speech=False), chunk_period_ms=1_000)
+
+    assert speech_unit is not None
+    assert speech_unit["is_speech"] is True
+    assert silence_unit is not None
+    assert silence_unit["is_speech"] is False
+
+
+def test_force_listen_marker_does_not_leak_across_irregular_chunk_boundaries():
+    buffer = MiniCPMO45PcmAppendBuffer()
+    forced = pcm_payload(15_000, speech=False)
+    forced["force_listen"] = True
+
+    assert buffer.append(forced, chunk_period_ms=1_000) is None
+    forced_unit = buffer.append(pcm_payload(2_000, speech=False), chunk_period_ms=1_000)
+    unforced_unit = buffer.append(pcm_payload(15_000, speech=False), chunk_period_ms=1_000)
+
+    assert forced_unit is not None
+    assert forced_unit["force_listen"] is True
+    assert unforced_unit is not None
+    assert unforced_unit["force_listen"] is False
+
+
+def test_pcm_append_rollback_restores_per_span_speech_markers():
+    buffer = MiniCPMO45PcmAppendBuffer()
+    assert buffer.append(pcm_payload(15_000), chunk_period_ms=1_000) is None
+    reservation = buffer.prepare_append(
+        pcm_payload(2_000, speech=False),
+        operation_id="mixed-unit",
+        chunk_period_ms=1_000,
+    )
+
+    assert reservation is not None
+    reservation.rollback()
+    first = buffer.append(pcm_payload(15_000, speech=False), chunk_period_ms=1_000)
+    second = buffer.flush(chunk_period_ms=1_000)
+
+    assert first is not None
+    assert first["is_speech"] is True
+    assert second is not None
+    assert second["is_speech"] is False
+
+
 def test_commit_without_speech_does_not_synthesize_terminal_audio():
     buffer = MiniCPMO45PcmAppendBuffer()
     buffer.append(pcm_payload(8_000, speech=False), chunk_period_ms=1_000)

tests/fullduplex/test_client.py

@@ -11,6 +11,9 @@ from vllm_omni.experimental.fullduplex.client import (
     read_pcm16_wav,
     write_pcm16_wav,
 )
+from vllm_omni.experimental.fullduplex.minicpmo45.policy import (
+    MiniCPMO45DuplexPolicy,
+)
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -32,6 +35,19 @@ def test_realtime_client_builds_explicit_native_duplex_url():
     }
 
 
+def test_seed_tts_initial_text_is_part_of_native_duplex_context():
+    prefix, suffix = MiniCPMO45DuplexPolicy.session_context_texts(
+        "Speak exactly.",
+        True,
+        "The quick brown fox.",
+    )
+
+    assert prefix == "<|im_start|>system\nSpeak exactly.\n<|audio_start|>"
+    assert suffix == (
+        "<|audio_end|><|im_end|>\n<|im_start|>user\nThe quick brown fox.<|im_end|>\n<|im_start|>assistant\n"
+    )
+
+
 def test_realtime_client_builds_resume_only_url_when_autostart_disabled():
     url = build_realtime_url(
         "ws://localhost:8099/v1/realtime?duplex=1",
@@ -86,6 +102,59 @@ async def test_realtime_client_configure_sends_explicit_ref_audio():
     assert session["ref_audio"] == "data:audio/wav;base64,AAAA"
 
 
+@pytest.mark.asyncio
+async def test_realtime_client_configure_sends_seed_tts_text_condition():
+    class Client(RealtimeDuplexClient):
+        def __init__(self):
+            super().__init__("ws://unused")
+            self.sent = []
+
+        async def send(self, event):
+            self.sent.append(event)
+            self.events.add({"type": "session.created"})
+
+    client = Client()
+
+    await client.configure(
+        "openbmb/MiniCPM-o-4_5",
+        instructions="Speak the requested text exactly.",
+        initial_user_text="The quick brown fox.",
+        timeout_s=1,
+    )
+
+    session = client.sent[0]["session"]
+    assert session["instructions"] == "Speak the requested text exactly."
+    assert session["extra_body"]["duplex_initial_user_text"] == "The quick brown fox."
+
+
+@pytest.mark.asyncio
+async def test_realtime_client_configure_explicit_tts_opts_out_of_native_duplex():
+    class Client(RealtimeDuplexClient):
+        def __init__(self):
+            super().__init__("ws://unused")
+            self.sent = []
+
+        async def send(self, event):
+            self.sent.append(event)
+            self.events.add({"type": "session.created"})
+
+    client = Client()
+
+    await client.configure(
+        "openbmb/MiniCPM-o-4_5",
+        native_duplex=False,
+        auto_response=False,
+        extra_body={"ref_audio": "data:audio/wav;base64,AAAA"},
+        timeout_s=1,
+    )
+
+    session_extra_body = client.sent[0]["session"]["extra_body"]
+    assert session_extra_body == {
+        "ref_audio": "data:audio/wav;base64,AAAA",
+        "minicpmo45_native_duplex": False,
+    }
+
+
 def test_realtime_event_collector_partitions_audio_by_response():
     collector = RealtimeEventCollector()
     collector.add({"type": "response.created", "response": {"id": "resp-a"}})
@@ -134,6 +203,26 @@ def test_realtime_event_collector_reports_engine_token_and_audio_intervals():
             },
             received_at_s=received_at_s,
         )
+    collector.add(
+        {
+            "type": "response.audio_transcript.delta",
+            "response_id": "resp-a",
+            "delta": "",
+        },
+        received_at_s=10.1,
+    )
+    collector.add(
+        {
+            "type": "response.audio_transcript.delta",
+            "response_id": "resp-a",
+            "delta": "hello",
+        },
+        received_at_s=10.15,
+    )
+    collector.add(
+        {"type": "response.done", "response": {"id": "resp-a"}},
+        received_at_s=10.4,
+    )
 
     timing = collector.timing_summary(
         after_s=10.0,
@@ -175,6 +264,19 @@ def test_realtime_event_collector_reports_engine_token_and_audio_intervals():
         },
         "max_chunk_gap_ms": 110.0,
     }
+    assert timing["request_metrics"] == {
+        "source": "client_monotonic_receive",
+        "measurement_origin": {
+            "ttft": "input_audio_buffer.commit client send to first non-empty text delta",
+            "ttfp": "input_audio_buffer.commit client send to first audio packet",
+            "rtf": "commit-to-last-audio receive time divided by emitted audio duration",
+        },
+        "ttft_ms": 250.0,
+        "ttfp_ms": 300.0,
+        "rtf": pytest.approx(1.916667),
+        "audio_generation_ms": 460.0,
+        "audio_duration_ms": 240.0,
+    }
 
 
 def test_response_timing_ignores_unowned_session_level_metrics():

tests/helpers/mark.py

@@ -2,10 +2,69 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Pytest marks and decorators for hardware / resource selection (CUDA, ROCm, …)."""
 
+from __future__ import annotations
+
+import re
+from functools import lru_cache
+from pathlib import Path
+
 import pytest
 from vllm.platforms import current_platform
 
-# Re-exported from tests.helpers.env (GPU wait + DeviceMemoryMonitor).
+# Marker description tag in ``pyproject.toml`` ``tool.pytest.ini_options.markers``.
+# Example: ``"H100: [hardware-resource] Tests that require H100 GPU"``.
+_HARDWARE_RESOURCE_MARKER_TAG = "[hardware-resource]"
+
+
+def _repo_root() -> Path:
+    return Path(__file__).resolve().parents[2]
+
+
+def _load_pyproject() -> dict:
+    path = _repo_root() / "pyproject.toml"
+    try:
+        try:
+            import tomllib
+        except ModuleNotFoundError:  # Python < 3.11
+            import tomli as tomllib  # type: ignore[no-redef]
+
+        with path.open("rb") as f:
+            return tomllib.load(f)
+    except Exception:
+        return {}
+
+
+@lru_cache(maxsize=1)
+def get_hardware_mark_list() -> frozenset[str]:
+    """Return hardware SKU marker names from ``pyproject.toml``.
+
+    A marker is treated as a hardware resource marker when its description
+    contains the ``[hardware-resource]`` tag, e.g.::
+
+        "H100: [hardware-resource] Tests that require H100 GPU"
+    """
+    data = _load_pyproject()
+    entries = data.get("tool", {}).get("pytest", {}).get("ini_options", {}).get("markers", [])
+    if not entries:
+        # Fallback when TOML parsing fails: scan tagged marker lines.
+        text = (_repo_root() / "pyproject.toml").read_text(encoding="utf-8")
+        return frozenset(
+            re.findall(
+                rf'^\s*"([A-Za-z0-9_]+)\s*:\s*{re.escape(_HARDWARE_RESOURCE_MARKER_TAG)}',
+                text,
+                flags=re.M,
+            )
+        )
+
+    names: set[str] = set()
+    for entry in entries:
+        text = str(entry)
+        if _HARDWARE_RESOURCE_MARKER_TAG not in text:
+            continue
+        name = text.split(":", 1)[0].strip()
+        if name:
+            names.add(name)
+    return frozenset(names)
 
 
 def cuda_marks(*, res: str, num_cards: int):

tests/helpers/minicpmo_4_5_duplex.py

@@ -0,0 +1,161 @@
+"""Shared fixtures and argument builders for MiniCPM-o 4.5 duplex E2E tests."""
+
+from __future__ import annotations
+
+import hashlib
+import uuid
+import wave
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+from huggingface_hub import snapshot_download
+
+from tests.helpers.runtime import OmniServerParams
+from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
+
+MODEL = "openbmb/MiniCPM-o-4_5"
+DEPLOY_CONFIG = modify_stage_config(
+    get_deploy_config_path("minicpmo_4_5_duplex.yaml"),
+    updates={
+        "base_config": get_deploy_config_path("minicpmo_4_5.yaml"),
+        # Talker context is 4096 (tts_config.max_position_embeddings); KV sizing
+        # is left automatic so duplex matches the simplex deploy profiles.
+        "stages": {
+            1: {"max_model_len": 4096},
+        },
+    },
+)
+ASSET_DIR = Path(__file__).resolve().parents[1] / "assets" / "minicpmo_4_5"
+RESPONSE_REQUIRED_WAV = ASSET_DIR / "response_required_16k.wav"
+RESPONSE_REQUIRED_SHA256 = "2e5fd4eb3ee434ce107ee3a0591fa624a33f7683c7462f45fe651c443c9af941"
+SOFT_INTERRUPT_WAV = ASSET_DIR / "soft_interrupt_16k.wav"
+SOFT_INTERRUPT_SHA256 = "cadae6d0ddc510310f16f8775d6379f30a5195369ccb54ff10a8e3f9a1f4a2ea"
+REF_AUDIO_RELATIVE_PATH = Path("assets") / "HT_ref_audio.wav"
+
+SERVER_PARAMS = [
+    pytest.param(
+        OmniServerParams(
+            model=MODEL,
+            stage_config_path=DEPLOY_CONFIG,
+            use_stage_cli=False,
+            server_args=["--trust-remote-code"],
+        ),
+        id="three-stage-single-gpu",
+    )
+]
+
+
+def validated_wav(path: Path, expected_sha256: str) -> Path:
+    actual_sha256 = hashlib.sha256(path.read_bytes()).hexdigest()
+    if actual_sha256 != expected_sha256:
+        raise AssertionError(
+            f"MiniCPM-o duplex fixture SHA256 mismatch: expected {expected_sha256}, got {actual_sha256}"
+        )
+    with wave.open(str(path), "rb") as wav_file:
+        actual_format = (
+            wav_file.getframerate(),
+            wav_file.getnchannels(),
+            wav_file.getsampwidth(),
+            wav_file.getcomptype(),
+        )
+    expected_format = (16_000, 1, 2, "NONE")
+    if actual_format != expected_format:
+        raise AssertionError(f"MiniCPM-o duplex fixture must be 16 kHz mono PCM16, got {actual_format}")
+    return path
+
+
+def validated_input_wav() -> Path:
+    return validated_wav(RESPONSE_REQUIRED_WAV, RESPONSE_REQUIRED_SHA256)
+
+
+def validated_soft_interrupt_wav() -> Path:
+    return validated_wav(SOFT_INTERRUPT_WAV, SOFT_INTERRUPT_SHA256)
+
+
+def resolve_ref_audio(model_prefix: str) -> Path:
+    # ``MODEL`` may be a local checkpoint directory instead of a Hub repo id
+    # (``Path("/prefix") / "/abs/path"`` collapses to the absolute path), so only
+    # fall back to the Hub cache when no such directory exists on disk.
+    model_root = Path(model_prefix) / MODEL if model_prefix else Path(MODEL)
+    if not model_root.is_dir():
+        model_root = Path(snapshot_download(MODEL, local_files_only=True))
+    ref_audio = model_root / REF_AUDIO_RELATIVE_PATH
+    if not ref_audio.is_file():
+        raise FileNotFoundError(f"MiniCPM-o checkpoint ref audio is missing: {ref_audio}")
+    return ref_audio
+
+
+def realtime_url(omni_server) -> str:
+    return f"ws://{omni_server.host}:{omni_server.port}/v1/realtime?duplex=1"
+
+
+def demo_args(
+    *,
+    omni_server,
+    input_wav: Path,
+    ref_audio: Path,
+    output_dir: Path,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        url=realtime_url(omni_server),
+        model=omni_server.model,
+        session_id=f"duplex-ci-single-{uuid.uuid4().hex}",
+        input_wav=str(input_wav),
+        ref_audio=str(ref_audio),
+        turn_input_wav=[],
+        output_dir=str(output_dir),
+        output_audio_format="pcm16",
+        chunk_ms=200,
+        realtime_input=True,
+        first_turn_ms=1400,
+        turn_duration_ms=[],
+        first_turn_transcript="duplex CI speech",
+        omit_transcript_hints=False,
+        validation_mode="response-required",
+        temperature=0.0,
+        scenario="sequential",
+        require_audio=True,
+        require_distinct_inputs=False,
+        expect_empty_turn=[],
+        short_ack_ms=350,
+        turns=1,
+        timeout_s=120.0,
+        model_policy_settle_ms=2000,
+    )
+
+
+def multi_session_args(
+    *,
+    omni_server,
+    input_wav: Path,
+    ref_audio: Path,
+    output_dir: Path,
+    response_required: bool,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        url=realtime_url(omni_server),
+        model=omni_server.model,
+        sessions=2,
+        input_wav=str(input_wav),
+        ref_audio=str(ref_audio),
+        session_input_wav=[],
+        session_expected_token=[],
+        turn_input_wav=[],
+        output_dir=str(output_dir),
+        realtime_input=True,
+        chunk_ms=200,
+        turns=1,
+        first_turn_ms=1400,
+        turn_duration_ms=[],
+        response_required=response_required,
+        temperature=0.0 if response_required else None,
+        disconnect_session_index=0,
+        takeover_session_index=1,
+        resume_after_ms=100,
+        expire_session_index=None,
+        expire_after_s=40.0,
+        verify_admission_limit=None,
+        model_policy_settle_ms=2000,
+        timeout_s=180.0,
+    )

tests/helpers/runtime.py

@@ -2881,7 +2881,12 @@ def iter_omni_server(
                 print("OmniServer stopping...")
         else:
             if stage_config_path is not None:
-                server_args += ["--stage-configs-path", stage_config_path]
+                from vllm_omni.entrypoints.utils import is_new_format_deploy_config
+
+                if is_new_format_deploy_config(stage_config_path):
+                    server_args += ["--deploy-config", stage_config_path]
+                else:
+                    server_args += ["--stage-configs-path", stage_config_path]
 
             with (
                 OmniServer(

tests/model_executor/models/minicpmo_4_5/test_code2wav_batching.py

@@ -19,10 +19,12 @@ class _FakeEncoder(nn.Module):
     def __init__(self):
         super().__init__()
         self.calls: list[int] = []
+        self.last_chunk_calls: list[bool] = []
 
     def forward_chunk(self, xs, last_chunk=False, cnn_cache=None, att_cache=None):
         batch, length, _ = xs.shape
         self.calls.append(batch)
+        self.last_chunk_calls.append(last_chunk)
         old_length = 0 if att_cache is None else att_cache.shape[3]
         output = xs[:, : max(1, length - 1)]
         cnn = xs[:, :1, :].transpose(1, 2).contiguous()
@@ -48,6 +50,8 @@ class _FakeEstimator(nn.Module):
         self.blocks = [_FakeBlock()]
         self.cfg_batches: list[int] = []
         self.speaker_order: list[list[float]] = []
+        self.register_buffer("att_cache_buffer", torch.ones(1), persistent=False)
+        self.register_buffer("cnn_cache_buffer", torch.ones(1), persistent=False)
 
     def t_embedder(self, time):
         return time[:, None]
@@ -77,6 +81,8 @@ class _FakeDecoder(nn.Module):
         self.estimator = _FakeEstimator()
         self.inference_cfg_rate = 0.7
         self.register_buffer("rand_noise", torch.zeros(1, 1, 100), persistent=False)
+        self.register_buffer("att_cache_buffer", torch.ones(1), persistent=False)
+        self.register_buffer("cnn_cache_buffer", torch.ones(1), persistent=False)
 
 
 class _FakeFlow(nn.Module):
@@ -156,6 +162,53 @@ def _model():
     return model, token2wav
 
 
+def test_adapter_releases_unused_upstream_streaming_buffers():
+    token2wav = _FakeToken2Wav()
+    decoder = token2wav.flow.decoder
+    modules = (decoder, decoder.estimator)
+
+    assert all(
+        getattr(module, name) is not None for module in modules for name in ("att_cache_buffer", "cnn_cache_buffer")
+    )
+
+    BatchedToken2Wav(token2wav)
+
+    assert all(
+        name in module._buffers and getattr(module, name) is None
+        for module in modules
+        for name in ("att_cache_buffer", "cnn_cache_buffer")
+    )
+
+
+def test_code2wav_resolves_hf_model_id_for_assets(mocker, tmp_path):
+    resolved_root = tmp_path / "snapshot"
+    resolved_root.mkdir()
+    config = _config()
+    config.model_config.model = "openbmb/MiniCPM-o-4_5"
+    config.model_config.revision = "test-revision"
+    config.model_config.stage_connector_config["extra"].pop("prompt_wav")
+    config.load_config = SimpleNamespace(download_dir="/model-cache")
+    model = MiniCPMO45Code2Wav(vllm_config=config)
+    mock_download = mocker.patch(
+        "vllm_omni.model_executor.model_loader.weight_utils.download_weights_from_hf_specific",
+        return_value=str(resolved_root),
+    )
+
+    assert model._resolve_model_root() == resolved_root
+    assert model.model_path == str(resolved_root)
+    assert model._default_prompt_wav == str(resolved_root / "assets" / "HT_ref_audio.wav")
+    mock_download.assert_called_once_with(
+        "openbmb/MiniCPM-o-4_5",
+        "/model-cache",
+        allow_patterns=[
+            "assets/HT_ref_audio.wav",
+            "assets/token2wav/*",
+        ],
+        revision="test-revision",
+        require_all=True,
+    )
+
+
 def _info(
     request_id: str,
     chunk_seq: int,
@@ -212,6 +265,18 @@ def test_adapter_runs_true_batch_cfg_and_splits_request_caches():
     assert cache1[0, 0, 0, 0, 0].item() == 20
 
 
+def test_fade_in_out_limits_overlap_to_available_previous_audio():
+    speech = torch.arange(6, dtype=torch.float32).reshape(1, -1)
+    previous = torch.full((1, 3), 2.0)
+    window = torch.hamming_window(8, periodic=False)
+
+    actual = BatchedToken2Wav._fade_in_out(speech, previous, window)
+
+    expected = speech.clone()
+    expected[..., :3] = speech[..., :3] * window[:3] + previous * window[-3:]
+    torch.testing.assert_close(actual, expected)
+
+
 def test_estimator_cache_stack_split_round_trip_preserves_cfg_rows():
     token2wav = _FakeToken2Wav()
     adapter = BatchedToken2Wav(token2wav)
@@ -259,6 +324,166 @@ def test_model_preserves_output_slots_and_prefers_runtime_codes():
     assert token2wav.flow.encoder.calls[-1] == 2
 
 
+def test_code2wav_projects_duplex_metadata_to_final_audio_output():
+    model, token2wav = _model()
+    segment = _info("duplex", 0, [10, 11])
+    segment_text_utf8 = torch.tensor(list(b"hello"), dtype=torch.uint8)
+    segment["meta"].update(
+        {
+            "duplex_epoch": 3,
+            "duplex_turn_id": 7,
+            "llm_output_text_utf8": segment_text_utf8,
+            "tts_is_last_chunk": True,
+            "turn_end": False,
+        }
+    )
+
+    segment_output = _forward(model, [segment])
+
+    assert segment_output.multimodal_outputs["meta.turn_end"][0].item() is False
+    # A Talker unit boundary only drains pending codec tokens. The official
+    # streaming path keeps Token2wav open until the assistant turn ends.
+    assert token2wav.flow.encoder.last_chunk_calls[-1] is False
+    assert "duplex" in model._states
+
+    final = _info("duplex", 1, [12, 13], last_chunk=True)
+    final["meta"].update(segment["meta"])
+    final["meta"]["chunk_seq"] = 1
+    final["meta"]["last_chunk"] = True
+    final["meta"]["turn_end"] = True
+    output = _forward(model, [final])
+
+    payload = output.multimodal_outputs
+    assert "meta" not in payload
+    assert payload["meta.duplex_epoch"][0].item() == 3
+    assert payload["meta.duplex_turn_id"][0].item() == 7
+    torch.testing.assert_close(
+        payload["meta.llm_output_text_utf8"][0],
+        segment_text_utf8,
+    )
+    assert payload["meta.tts_is_last_chunk"][0].item() is True
+    assert payload["meta.turn_end"][0].item() is True
+    assert token2wav.flow.encoder.last_chunk_calls[-1] is True
+    assert "duplex" not in model._states
+
+
+def test_initial_empty_segment_marker_initializes_stream_without_audio():
+    model, token2wav = _model()
+    boundary = _info("duplex", 0, [])
+    boundary["meta"].update(
+        {
+            "code_flat_numel": 0,
+            "tts_is_last_chunk": True,
+            "turn_end": False,
+        }
+    )
+
+    output = _forward(model, [boundary])
+
+    assert output.multimodal_outputs["model_outputs"][0].numel() == 0
+    assert "duplex" in model._states
+    assert token2wav.hift.calls == []
+
+    resumed = _info(
+        "duplex",
+        1,
+        [4218, 4218, 4218, 10, 11, 12, 13, 14],
+    )
+    output = _forward(model, [resumed])
+
+    assert output.multimodal_outputs["model_outputs"][0].numel() > 0
+    assert "duplex" in model._states
+
+
+def test_shared_runtime_prompt_recreates_missing_file_before_second_owner(tmp_path, monkeypatch):
+    monkeypatch.setattr("tempfile.gettempdir", lambda: str(tmp_path))
+    model, _ = _model()
+    reference = torch.tensor([0.0, 0.25, -0.25, 0.0])
+
+    first = _info("voice-a", 0, [10, 11])
+    first["codes"]["ref"] = reference
+    first["meta"]["ref_audio_sr"] = 16000
+    first["meta"].pop("prompt_cache_id")
+    _forward(model, [first], request_ids=["internal-a"])
+
+    prompt_key = model._request_prompt_keys["internal-a"]
+    prompt_path = Path(model._runtime_prompts[prompt_key].path)
+    prompt_path.unlink()
+
+    second = _info("voice-b", 0, [12, 13])
+    second["codes"]["ref"] = reference
+    second["meta"]["ref_audio_sr"] = 16000
+    second["meta"].pop("prompt_cache_id")
+    _forward(model, [second], request_ids=["internal-b"])
+
+    assert prompt_path.is_file()
+    assert model._runtime_prompts[prompt_key].owners == {"internal-a", "internal-b"}
+
+    model.on_requests_finished(["internal-a"])
+    assert prompt_path.is_file()
+    assert model._runtime_prompts[prompt_key].owners == {"internal-b"}
+
+    model.on_requests_finished(["internal-b"])
+    assert not prompt_path.exists()
+    assert prompt_key not in model._runtime_prompts
+
+
+def test_runtime_prompt_write_failure_does_not_publish_partial_file(tmp_path, monkeypatch):
+    monkeypatch.setattr("tempfile.gettempdir", lambda: str(tmp_path))
+    model, _ = _model()
+    reference = torch.tensor([0.0, 0.25, -0.25, 0.0])
+
+    def fail_after_partial_write(path, *_args, **_kwargs):
+        Path(path).write_bytes(b"partial")
+        raise OSError("simulated write failure")
+
+    monkeypatch.setattr(
+        "vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_code2wav.sf.write",
+        fail_after_partial_write,
+    )
+
+    with pytest.raises(OSError, match="simulated write failure"):
+        model._materialize_runtime_prompt(reference, 16000)
+
+    assert len(model._runtime_prompts) == 1
+    entry = next(iter(model._runtime_prompts.values()))
+    assert not Path(entry.path).exists()
+    assert list(Path(entry.path).parent.iterdir()) == []
+
+
+def test_runtime_prompt_files_are_isolated_between_model_instances(tmp_path, monkeypatch):
+    monkeypatch.setattr("tempfile.gettempdir", lambda: str(tmp_path))
+    first_model, _ = _model()
+    second_model, _ = _model()
+    reference = torch.tensor([0.0, 0.25, -0.25, 0.0])
+
+    def runtime_ref_info(request_id: str):
+        info = _info(request_id, 0, [10, 11])
+        info["codes"]["ref"] = reference
+        info["meta"]["ref_audio_sr"] = 16000
+        info["meta"].pop("prompt_cache_id")
+        return info
+
+    _forward(first_model, [runtime_ref_info("voice-a")], request_ids=["internal-a"])
+    _forward(second_model, [runtime_ref_info("voice-b")], request_ids=["internal-b"])
+
+    first_key = first_model._request_prompt_keys["internal-a"]
+    second_key = second_model._request_prompt_keys["internal-b"]
+    first_path = Path(first_model._runtime_prompts[first_key].path)
+    second_path = Path(second_model._runtime_prompts[second_key].path)
+    assert first_key == second_key
+    assert first_path != second_path
+    assert first_path.is_file()
+    assert second_path.is_file()
+
+    first_model.on_requests_finished(["internal-a"])
+    assert not first_path.exists()
+    assert second_path.is_file()
+
+    second_model.on_requests_finished(["internal-b"])
+    assert not second_path.exists()
+
+
 def test_mixed_final_exact_buckets_keep_order_and_release_only_final_states():
     model, _ = _model()
     _forward(
@@ -485,27 +710,37 @@ def test_reference_voice_and_duplex_metadata_follow_request_lifecycle():
     model, _ = _model()
     first = _info("voice-a", 0, [1, 2])
     first["codes"]["ref"] = torch.linspace(-0.1, 0.1, 160)
+    segment_text_utf8 = torch.tensor(list(b"hello"), dtype=torch.uint8)
     first["meta"].update(
         ref_audio_sr=16000,
-        native_duplex_segment_text="hello",
+        llm_output_text_utf8=segment_text_utf8,
         duplex_turn_id=7,
         duplex_epoch=3,
     )
     first["meta"].pop("prompt_cache_id")
 
     output = _forward(model, [first])
-    prompt_cache_id, prompt_wav = model._owned_prompt_wavs["voice-a"]
-    assert prompt_cache_id.startswith("voice-a:")
+    prompt_key = model._request_prompt_keys["voice-a"]
+    prompt = model._runtime_prompts[prompt_key]
+    prompt_cache_id, prompt_wav = prompt.cache_id, prompt.path
+    assert prompt_cache_id.startswith("runtime-ref-")
     assert Path(prompt_wav).is_file()
-    assert bytes(output.multimodal_outputs["meta.llm_output_text_utf8"][0].tolist()).decode() == "hello"
+    torch.testing.assert_close(
+        output.multimodal_outputs["meta.llm_output_text_utf8"][0],
+        segment_text_utf8,
+    )
     assert output.multimodal_outputs["meta.duplex_turn_id"][0].item() == 7
     assert output.multimodal_outputs["meta.duplex_epoch"][0].item() == 3
 
     final = _info("voice-a", 1, [3, 4], last_chunk=True)
     final["meta"].pop("prompt_cache_id")
+    final["meta"]["tts_is_last_chunk"] = True
     output = _forward(model, [final])
 
     assert output.multimodal_outputs["meta.tts_is_last_chunk"][0].item() is True
-    assert "voice-a" not in model._owned_prompt_wavs
+    assert model._request_prompt_keys["voice-a"] == prompt_key
+    model.on_requests_finished(["voice-a"])
+    assert "voice-a" not in model._request_prompt_keys
+    assert prompt_key not in model._runtime_prompts
     assert not Path(prompt_wav).exists()
     assert (prompt_cache_id, prompt_wav) not in model.backend._prompt_features

tests/model_executor/models/minicpmo_4_5/test_pipeline.py

@@ -24,8 +24,6 @@ from vllm_omni.config.pipeline_registry import OMNI_PIPELINES
 from vllm_omni.config.stage_config import (
     PipelineConfig,
     StageExecutionType,
-    load_deploy_config,
-    merge_pipeline_deploy,
 )
 from vllm_omni.model_executor.models.registry import _OMNI_MODELS
 
@@ -36,7 +34,6 @@ _PIPELINE_KEY = "minicpmo_4_5"
 _DEPLOY_DIR = Path(__file__).resolve().parents[4] / "vllm_omni" / "deploy"
 _DEPLOY_LAYOUTS = {
     "minicpmo_4_5.yaml": ["0", "0", "0"],
-    "minicpmo_4_5_batching.yaml": ["0", "1", "1"],
     "minicpmo_4_5_2gpu.yaml": ["0", "1", "1"],
     "minicpmo_4_5_3gpu.yaml": ["0", "1", "2"],
     "minicpmo_4_5_8x4090.yaml": ["0,1,2,3", "4", "5"],
@@ -138,45 +135,6 @@ class TestPipelineTopology:
 
 
 class TestDeployTopology:
-    @pytest.mark.parametrize(("filename", "devices"), _DEPLOY_LAYOUTS.items())
-    def test_deploy_resolves_three_stage_pipeline(self, filename: str, devices: list[str]) -> None:
-        deploy = load_deploy_config(_DEPLOY_DIR / filename)
-        pipeline = OMNI_PIPELINES[deploy.pipeline]
-        stages = merge_pipeline_deploy(pipeline, deploy)
-
-        assert deploy.pipeline == _PIPELINE_KEY
-        assert deploy.async_chunk is True
-        assert [stage.stage_id for stage in stages] == [0, 1, 2]
-        assert [stage.yaml_runtime["devices"] for stage in stages] == devices
-        assert "additional_config" not in stages[1].yaml_engine_args
-        assert "skip_mm_profiling" not in stages[0].yaml_engine_args
-        assert stages[1].yaml_engine_args["skip_mm_profiling"] is True
-        assert stages[2].yaml_engine_args["skip_mm_profiling"] is True
-        assert stages[1].yaml_extras["output_connectors"]["to_stage_2"] == "connector_of_shared_memory"
-        assert stages[2].yaml_extras["input_connectors"]["from_stage_1"] == "connector_of_shared_memory"
-        connector = deploy.connectors["connector_of_shared_memory"]
-        assert connector["name"] == "SharedMemoryConnector"
-        assert connector["extra"]["codec_chunk_frames"] == 25
-        assert connector["extra"]["codec_left_context_frames"] == 3
-        assert connector["extra"]["connector_get_max_wait_first_chunk"] == 3000
-        assert connector["extra"]["connector_get_max_wait"] == 300
-        expected_processor = "tts2code2wav_async_chunk" if deploy.async_chunk else "tts2code2wav_full_payload"
-        assert stages[1].yaml_engine_args["custom_process_next_stage_input_func"].endswith(expected_processor)
-        if filename == "minicpmo_4_5.yaml":
-            assert [stage.yaml_engine_args["max_num_seqs"] for stage in stages] == [4, 4, 4]
-            assert [stage.yaml_engine_args["gpu_memory_utilization"] for stage in stages] == [
-                0.55,
-                0.22,
-                0.22,
-            ]
-            assert stages[0].yaml_engine_args["limit_mm_per_prompt"] == {"video": {"count": 1, "num_frames": 32}}
-        elif filename in {"minicpmo_4_5_batching.yaml", "minicpmo_4_5_2gpu.yaml"}:
-            assert [stage.yaml_engine_args["gpu_memory_utilization"] for stage in stages] == [
-                0.9,
-                0.55,
-                0.35,
-            ]
-
     def test_pipeline_exposes_full_and_async_payload_hooks(self) -> None:
         pipeline = OMNI_PIPELINES[_PIPELINE_KEY]
         talker = pipeline.get_stage(1)

tests/model_executor/models/minicpmo_4_5/test_talker_batching.py

@@ -14,6 +14,7 @@ from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni import (
 from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts import (
     MiniCPMO45OmniTTSForConditionalGeneration,
     _max_audio_tokens,
+    _restore_weight_norm_weight,
 )
 from vllm_omni.utils.mm_outputs import to_payload_element
 
@@ -56,6 +57,8 @@ def _make_talker() -> MiniCPMO45OmniTTSForConditionalGeneration:
     talker._request_generators = {}
     talker._request_audio_states = {}
     talker._deferred_cleanup_ids = set()
+    talker._codec_min_tokens = 50
+    talker._codec_seed = 42
     return talker
 
 
@@ -81,6 +84,23 @@ def test_audio_token_limit_scales_with_condition_length(
     assert _max_audio_tokens(condition_tokens) == expected
 
 
+def test_weight_norm_restore_matches_checkpoint_parametrization_in_bfloat16() -> None:
+    generator = torch.Generator().manual_seed(42)
+    weight_v = torch.randn(8, 16, generator=generator, dtype=torch.bfloat16)
+    weight_g = torch.rand(8, 1, generator=generator, dtype=torch.bfloat16)
+    linear = nn.utils.parametrizations.weight_norm(
+        nn.Linear(16, 8, bias=False, dtype=torch.bfloat16),
+        dim=0,
+    )
+    with torch.no_grad():
+        linear.parametrizations.weight.original0.copy_(weight_g)
+        linear.parametrizations.weight.original1.copy_(weight_v)
+
+    restored = _restore_weight_norm_weight(weight_g, weight_v)
+
+    assert torch.equal(restored, linear.weight)
+
+
 def test_talker_emits_request_aligned_codec_deltas_after_compaction(mocker) -> None:
     talker = _make_talker()
     seen: list[tuple[str, list[float], list[int]]] = []
@@ -114,9 +134,73 @@ def test_talker_emits_request_aligned_codec_deltas_after_compaction(mocker) -> N
     assert _routed(output, 0)["codes"]["audio"].tolist() == [[2]]
     assert _routed(output, 1)["codes"]["audio"].tolist() == [[3]]
     assert _routed(output, 0)["meta"]["finished"].item() is False
+    assert set(output.multimodal_outputs["meta"]) == {"finished"}
     assert talker.compute_logits(output.text_hidden_states).argmax(dim=-1).tolist() == [0, 0]
 
 
+def test_talker_projects_request_aligned_duplex_metadata(mocker) -> None:
+    talker = _make_talker()
+    mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(2))
+    infos = [
+        {
+            "request_id": "req-a",
+            "native_duplex": True,
+            "duplex": {"epoch": 3, "turn_id": 7},
+            "ids": {"tts": [41]},
+            "meta": {
+                "native_duplex_segment_text": "first",
+                "turn_eos_token_id": 99,
+            },
+            "audio_codes": {"accumulated": torch.empty(0, dtype=torch.long)},
+        },
+        {
+            "request_id": "req-b",
+            "native_duplex": True,
+            "duplex": {"epoch": 4, "turn_id": 8},
+            "ids": {"tts": [42, 99]},
+            "meta": {
+                "native_duplex_segment_text": "second",
+                "turn_eos_token_id": 99,
+            },
+            "audio_codes": {"accumulated": torch.empty(0, dtype=torch.long)},
+        },
+    ]
+
+    output = talker.make_omni_output(
+        torch.ones(2, 2),
+        model_intermediate_buffer=infos,
+        request_token_spans=[(0, 1), (1, 2)],
+    )
+
+    meta = output.multimodal_outputs["meta"]
+    assert [value.item() for value in meta["native_duplex"]] == [True, True]
+    assert [value.item() for value in meta["duplex_epoch"]] == [3, 4]
+    assert [value.item() for value in meta["duplex_turn_id"]] == [7, 8]
+    assert "native_duplex_segment_text" not in meta
+    assert [bytes(value.tolist()).decode("utf-8") for value in meta["llm_output_text_utf8"]] == [
+        "first",
+        "second",
+    ]
+    assert [value.item() for value in meta["turn_end"]] == [False, True]
+
+
+def test_talker_rejects_native_duplex_without_fence_identity(mocker) -> None:
+    talker = _make_talker()
+    mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(2))
+    info = {
+        "request_id": "req-missing-fence",
+        "native_duplex": True,
+        "audio_codes": {"accumulated": torch.empty(0, dtype=torch.long)},
+    }
+
+    with pytest.raises(RuntimeError, match="requires non-negative integer epoch and turn_id"):
+        talker.make_omni_output(
+            torch.ones(1, 2),
+            model_intermediate_buffer=[info],
+            request_token_spans=[(0, 1)],
+        )
+
+
 def test_incomplete_prefill_emits_no_code_and_does_not_advance_state(mocker) -> None:
     talker = _make_talker()
     sample = mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(2))
@@ -179,7 +263,7 @@ def test_eos_is_terminal_once_and_never_enters_codec_history(mocker) -> None:
     assert talker.compute_logits(second.text_hidden_states).argmax(dim=-1).tolist() == [1]
 
 
-def test_max_token_terminal_includes_only_new_codec_delta(mocker) -> None:
+def test_max_token_terminal_drops_unconsumed_codec_delta(mocker) -> None:
     talker = _make_talker()
     mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(3))
     info = {
@@ -194,8 +278,10 @@ def test_max_token_terminal_includes_only_new_codec_delta(mocker) -> None:
         request_token_spans=[(0, 1)],
     )
 
-    assert info["audio_codes"]["accumulated"].tolist() == [4, 5, 3]
-    assert output.multimodal_outputs["codes"]["audio"][0].tolist() == [[3]]
+    # MiniCPMTTS.generate_chunk samples once at the max-token boundary to
+    # advance RNG state, but the sampled code is not fed into KV or returned.
+    assert info["audio_codes"]["accumulated"].tolist() == [4, 5]
+    assert output.multimodal_outputs["codes"]["audio"][0].shape == (0, 1)
     assert output.multimodal_outputs["meta"]["finished"][0].item() is True
     assert talker.compute_logits(output.text_hidden_states).argmax(dim=-1).tolist() == [1]
 
@@ -268,6 +354,94 @@ def test_empty_speech_segment_finishes_without_sampling_codes() -> None:
     assert decode_embeds.shape == (1, 4)
 
 
+def test_chunked_prefill_tail_aligns_condition_with_prompt_length(mocker) -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(1, 2)
+    condition = torch.arange(18, dtype=torch.float32).reshape(9, 2)
+    mocker.patch.object(talker, "_build_condition_embeddings", return_value=condition)
+
+    _, embeds, _ = talker.preprocess(
+        torch.zeros(9, dtype=torch.long),
+        None,
+        _omni_is_prefill=True,
+        _omni_num_computed_tokens=59,
+        _omni_prompt_len=68,
+        request_id="req-chunked-prefill",
+        tts_token_ids=torch.tensor([1]),
+        tts_hidden_states=torch.ones(1, 2),
+    )
+
+    assert torch.equal(embeds, condition)
+    state = talker._request_audio_states["req-chunked-prefill"]
+    assert state["min_tokens"] == 50
+    assert state["max_tokens"] == 64
+
+
+@pytest.mark.parametrize(
+    ("meta", "expected_min_tokens"),
+    [
+        ({"turn_start": True}, 0),
+        ({}, 26),
+        ({"turn_end": True}, 0),
+    ],
+)
+def test_native_duplex_prefill_uses_official_chunk_limits(
+    mocker,
+    meta,
+    expected_min_tokens,
+) -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(1, 2)
+    mocker.patch.object(
+        talker,
+        "_build_condition_embeddings",
+        return_value=torch.ones(3, 2),
+    )
+
+    talker.preprocess(
+        torch.zeros(3, dtype=torch.long),
+        None,
+        _omni_is_prefill=True,
+        request_id="req-duplex-chunk",
+        native_duplex=True,
+        meta=meta,
+        tts_token_ids=torch.tensor([1]),
+        tts_hidden_states=torch.ones(1, 2),
+    )
+
+    state = talker._request_audio_states["req-duplex-chunk"]
+    assert state["min_tokens"] == expected_min_tokens
+    assert state["max_tokens"] == 26
+
+
+def test_native_duplex_condition_matches_official_text_plus_audio_bos() -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(16, 2)
+    talker.projector_semantic = nn.Identity()
+    talker._normalize = False
+    talker._text_eos_id = 14
+    talker._tts_bos_id = 15
+    with torch.no_grad():
+        talker.emb_text.weight.copy_(torch.arange(32, dtype=torch.float32).reshape(16, 2))
+
+    token_ids = torch.tensor([2, 3])
+    hidden_states = torch.tensor([[0.5, 1.0], [1.5, 2.0]])
+
+    condition = talker._build_condition_embeddings(
+        token_ids,
+        hidden_states,
+        native_duplex=True,
+    )
+
+    expected_text = talker.emb_text(token_ids) + hidden_states
+    expected = torch.cat(
+        [expected_text, talker.emb_text(torch.tensor([talker._tts_bos_id]))],
+        dim=0,
+    )
+    assert torch.equal(condition, expected)
+    assert condition.shape[0] == token_ids.shape[0] + 1
+
+
 def test_request_cleanup_evicts_ar_rng_and_decode_state() -> None:
     talker = _make_talker()
     talker._request_generators["req-done"] = torch.Generator()

tests/model_executor/stage_input_processors/test_minicpmo_4_5_async_chunk.py

@@ -43,6 +43,27 @@ def _delta(*codes: int):
     }
 
 
+def _duplex_delta(
+    *codes: int,
+    epoch: int = 3,
+    turn_id: int = 7,
+    text: str = "segment",
+    turn_end: bool = False,
+):
+    text_utf8 = torch.tensor(list(text.encode("utf-8")), dtype=torch.uint8)
+    return {
+        "codes": {"audio": torch.tensor(codes, dtype=torch.long).reshape(-1, 1)},
+        "meta": {
+            "finished": torch.tensor(False),
+            "native_duplex": torch.tensor(True),
+            "duplex_epoch": torch.tensor(epoch),
+            "duplex_turn_id": torch.tensor(turn_id),
+            "llm_output_text_utf8": text_utf8,
+            "turn_end": torch.tensor(turn_end),
+        },
+    }
+
+
 def _codes(payload) -> list[int]:
     assert payload.codes is not None
     assert isinstance(payload.codes.audio, torch.Tensor)
@@ -108,29 +129,57 @@ def test_short_final_flushes_silence_prefix_and_tail() -> None:
     assert final.meta.finished.item() is True
 
 
+def test_duplex_turn_end_waits_for_terminal_codec_flush() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    body = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(25), turn_end=True),
+        request,
+        False,
+    )
+    final = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(turn_end=True),
+        request,
+        True,
+    )
+
+    assert body is not None
+    assert body.meta.last_chunk is False
+    assert body.meta.turn_end is False
+    assert final is not None
+    assert _codes(final) == [22, 23, 24]
+    assert final.meta.last_chunk is True
+    assert final.meta.turn_end is True
+
+
 def test_first_chunk_forwards_reference_voice_and_duplex_identity() -> None:
     manager = _manager()
     request = _request("req")
     request.additional_information = {
         "codes": {"ref": [0.1, -0.1]},
-        "meta": {
-            "ref_audio_sr": 16000,
-            "native_duplex_segment_text": "hello",
-            "segment_end": True,
-            "turn_end": True,
-        },
-        "duplex": {"epoch": 3, "model_turn_id": 7},
+        "meta": {"ref_audio_sr": 16000},
     }
 
-    payload = tts2code2wav_async_chunk(manager, _delta(*range(7)), request, True)
+    payload = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(7), text="hello", turn_end=True),
+        request,
+        True,
+    )
 
     assert payload is not None
     assert payload.codes.ref.tolist() == pytest.approx([0.1, -0.1])
     assert payload.meta.ref_audio_sr == 16000
-    assert payload.meta.native_duplex_segment_text == "hello"
+    torch.testing.assert_close(
+        payload.meta.llm_output_text_utf8,
+        torch.tensor(list(b"hello"), dtype=torch.uint8),
+    )
     assert payload.meta.duplex_epoch == 3
     assert payload.meta.duplex_turn_id == 7
-    assert payload.meta.segment_end is True
+    assert payload.meta.tts_is_last_chunk is True
     assert payload.meta.turn_end is True
 
 
@@ -211,6 +260,186 @@ def test_empty_final_releases_wait_gate_once() -> None:
     assert duplicate is None
 
 
+def test_empty_duplex_boundary_uses_zero_length_transport_placeholder() -> None:
+    manager = _manager()
+
+    boundary = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(text="boundary"),
+        _request("req-duplex"),
+        True,
+    )
+
+    assert boundary is not None
+    assert _codes(boundary) == [0]
+    assert boundary.meta.code_flat_numel == 0
+    assert boundary.meta.last_chunk is False
+    assert boundary.meta.is_segment_finished.item() is False
+    torch.testing.assert_close(
+        boundary.meta.llm_output_text_utf8,
+        torch.tensor(list(b"boundary"), dtype=torch.uint8),
+    )
+
+
+def test_duplex_segments_preserve_stream_state_without_closing_turn() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+    request.additional_information = {
+        "codes": {"ref": [0.1, 0.2, 0.3]},
+        "meta": {"ref_audio_sr": 16000},
+    }
+
+    first = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(10, 11, text="first"),
+        request,
+        True,
+    )
+    second = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(12, 13, text="second"),
+        request,
+        True,
+    )
+
+    assert first is not None
+    assert second is not None
+    assert first.meta.last_chunk is False
+    assert second.meta.last_chunk is False
+    assert first.codes is not None
+    assert torch.allclose(first.codes.ref, torch.tensor([0.1, 0.2, 0.3]))
+    assert first.meta.ref_audio_sr == 16000
+    assert second.codes is not None
+    assert second.codes.ref is None
+    assert second.meta.cache_epoch == first.meta.cache_epoch
+    assert second.meta.chunk_seq == first.meta.chunk_seq + 1
+    assert second.meta.duplex_epoch == 3
+    assert second.meta.duplex_turn_id == 7
+    torch.testing.assert_close(
+        second.meta.llm_output_text_utf8,
+        torch.tensor(list(b"second"), dtype=torch.uint8),
+    )
+    assert second.meta.tts_is_last_chunk is True
+    assert second.meta.turn_end is False
+    assert first.meta.is_segment_finished.item() is False
+    assert second.meta.is_segment_finished.item() is False
+
+
+def test_duplex_short_units_wait_for_minimum_stream_body() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    first = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(10, 11, 12, text="first"),
+        request,
+        True,
+    )
+    second = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(13, 14, text="first"),
+        request,
+        True,
+    )
+
+    assert first is not None
+    assert _codes(first) == [0]
+    assert first.meta.code_flat_numel == 0
+    assert first.meta.last_chunk is False
+    assert first.meta.tts_is_last_chunk is True
+    assert second is not None
+    assert _codes(second) == [4218, 4218, 4218, 10, 11, 12, 13, 14]
+    assert second.meta.code_flat_numel == 8
+    torch.testing.assert_close(
+        second.meta.llm_output_text_utf8,
+        torch.tensor(list(b"firstfirst"), dtype=torch.uint8),
+    )
+
+
+def test_duplex_empty_finish_callback_does_not_replay_previous_text() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    first = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(25), text="first"),
+        request,
+        False,
+    )
+    boundary = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(text="first"),
+        request,
+        True,
+    )
+    second = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(25, 50), text="second"),
+        request,
+        False,
+    )
+
+    assert first is not None
+    assert boundary is not None
+    assert boundary.meta.code_flat_numel == 0
+    assert second is not None
+    torch.testing.assert_close(
+        second.meta.llm_output_text_utf8,
+        torch.tensor(list(b"second"), dtype=torch.uint8),
+    )
+
+
+def test_duplex_short_tail_does_not_replay_previous_segment_text() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    def chunk(codes, text: str, finished: bool):
+        return tts2code2wav_async_chunk(
+            manager,
+            _duplex_delta(*codes, text=text),
+            request,
+            finished,
+        )
+
+    first = chunk(range(25), "和上海之间", False)
+    first_tail = chunk([25, 26], "和上海之间", True)
+    assert first is not None
+    assert first.meta.llm_output_text_utf8.tolist() == list("和上海之间".encode())
+    assert first_tail is not None
+    assert first_tail.meta.code_flat_numel == 0
+    assert chunk([27, 28, 29], "的距离大约是", False) is None
+    next_flush = chunk([], "的距离大约是", True)
+    assert next_flush is not None
+    assert next_flush.meta.llm_output_text_utf8.tolist() == list("的距离大约是".encode())
+
+
+def test_duplex_turn_end_closes_epoch_and_next_turn_restarts_sequence() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    turn_end = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(14, turn_id=7, turn_end=True),
+        request,
+        True,
+    )
+    next_turn = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(20, turn_id=8),
+        request,
+        True,
+    )
+
+    assert turn_end is not None
+    assert next_turn is not None
+    assert turn_end.meta.last_chunk is True
+    assert turn_end.meta.turn_end is True
+    assert turn_end.meta.is_segment_finished.item() is True
+    assert next_turn.meta.cache_epoch == turn_end.meta.cache_epoch + 1
+    assert next_turn.meta.chunk_seq == 0
+    assert next_turn.meta.last_chunk is False
+
+
 def test_staggered_requests_keep_accumulators_isolated() -> None:
     manager = _manager()
     req_a = _request("a")

tests/model_executor/stage_input_processors/test_minicpmo_4_5_omni.py

@@ -128,13 +128,125 @@ def test_native_duplex_speak_segment_reaches_split_talker() -> None:
     info = converted["model_intermediate_buffer"]
     assert info["native_duplex"] is True
     assert info["ids"]["tts"] == [21, 22]
-    assert converted["prompt_token_ids"] == [0, 0, 0, 0]
+    assert converted["prompt_token_ids"] == [0, 0, 0]
     assert info["meta"]["replace_streaming_prompt"] is True
+    assert info["meta"]["next_stage_prompt_len"] == 3
+    assert info["meta"]["turn_start"] is True
     assert info["meta"]["segment_end"] is True
     assert info["duplex"]["epoch"] == 3
     assert info["duplex"]["turn_id"] == 7
 
 
+def test_native_duplex_continuation_appends_only_new_talker_condition() -> None:
+    prompt_ids = [101, 102]
+    token_ids = {
+        "tts_bos_token_id": 9301,
+        "tts_eos_token_id": 9302,
+        "listen_token_id": 9303,
+        "speak_token_id": 9304,
+        "chunk_eos_token_id": 9308,
+        "chunk_tts_eos_token_id": 9309,
+        "turn_eos_token_id": 9310,
+    }
+    context = SimpleNamespace(
+        bridge_states={
+            "duplex": {
+                "epoch": 3,
+                "model_turn_id": 7,
+            }
+        }
+    )
+
+    first_ids = [9304, 21, 22, 9308]
+    first = _output(
+        prompt_ids=prompt_ids,
+        output_ids=first_ids,
+        latent=torch.arange(24, dtype=torch.float32).reshape(6, 4),
+        multimodal_output={
+            "duplex_prompt_token_ids": prompt_ids,
+            "meta": token_ids,
+        },
+    )
+    second_ids = [*first_ids, 9304, 23, 24, 9308]
+    second = _output(
+        prompt_ids=prompt_ids,
+        output_ids=second_ids,
+        latent=torch.arange(40, dtype=torch.float32).reshape(10, 4),
+        multimodal_output={
+            "duplex_prompt_token_ids": prompt_ids,
+            "meta": token_ids,
+        },
+    )
+    third_ids = [*second_ids, 9304, 25, 26, 9308]
+    third = _output(
+        prompt_ids=prompt_ids,
+        output_ids=third_ids,
+        latent=torch.arange(56, dtype=torch.float32).reshape(14, 4),
+        multimodal_output={
+            "duplex_prompt_token_ids": prompt_ids,
+            "meta": token_ids,
+        },
+    )
+
+    first_input = llm2tts([first], prompt=[{}], _streaming_context=context)[0]
+    second_input = llm2tts([second], prompt=[{}], _streaming_context=context)[0]
+    context.bridge_states["duplex"]["model_turn_id"] = 8
+    third_input = llm2tts([third], prompt=[{}], _streaming_context=context)[0]
+
+    assert first_input["model_intermediate_buffer"]["ids"]["tts"] == [21, 22]
+    assert second_input["model_intermediate_buffer"]["ids"]["tts"] == [23, 24]
+    assert third_input["model_intermediate_buffer"]["ids"]["tts"] == [25, 26]
+    assert first_input["model_intermediate_buffer"]["meta"]["turn_start"] is True
+    assert second_input["model_intermediate_buffer"]["meta"]["turn_start"] is False
+    assert third_input["model_intermediate_buffer"]["meta"]["turn_start"] is True
+    assert first_input["model_intermediate_buffer"]["meta"]["replace_streaming_prompt"] is True
+    assert "replace_streaming_prompt" not in second_input["model_intermediate_buffer"]["meta"]
+    assert third_input["model_intermediate_buffer"]["meta"]["replace_streaming_prompt"] is True
+    assert second_input["model_intermediate_buffer"]["meta"]["next_stage_prompt_len"] == 3
+    assert second_input["prompt_token_ids"] == [0, 0, 0]
+
+
+def test_native_duplex_transcript_decodes_the_talker_condition_slice() -> None:
+    prompt_ids = [101, 102]
+    metadata = {
+        "tts_bos_token_id": 9301,
+        "tts_eos_token_id": 9302,
+        "listen_token_id": 9303,
+        "speak_token_id": 9304,
+        "chunk_eos_token_id": 9308,
+        "chunk_tts_eos_token_id": 9309,
+        "turn_eos_token_id": 9310,
+    }
+    token_text = {21: "杭州", 22: "和", 23: "州和", 24: "上海之间", 25: "大"}
+
+    def output(token_ids: list[int]):
+        return _output(
+            prompt_ids=prompt_ids,
+            output_ids=token_ids,
+            latent=torch.zeros((len(prompt_ids) + len(token_ids), 1)),
+            multimodal_output={
+                "duplex_prompt_token_ids": prompt_ids,
+                "meta": metadata,
+            },
+        )
+
+    context = SimpleNamespace(
+        bridge_states={"duplex": {"epoch": 3, "model_turn_id": 7}},
+        source_token_decoder=lambda ids, **_: "".join(token_text.get(int(token_id), "") for token_id in ids),
+    )
+    first_ids = [9304, 21, 22, 9308]
+    first_info = llm2tts([output(first_ids)], prompt=[{}], _streaming_context=context)[0]["model_intermediate_buffer"]
+    second_info = llm2tts(
+        [output([*first_ids, 23, 24, 25, 9308])],
+        prompt=[{}],
+        _streaming_context=context,
+    )[0]["model_intermediate_buffer"]
+
+    assert first_info["meta"]["native_duplex_segment_text"] == "杭州和"
+    assert second_info["ids"]["tts"] == [24, 25]
+    assert second_info["meta"]["native_duplex_segment_text"] == "上海之间大"
+
+
 def test_native_duplex_requires_tokenizer_boundary_metadata() -> None:
     latent = torch.zeros((3, 4))
     source = _output(

tests/model_executor/stage_input_processors/test_qwen3_omni_streaming_helpers.py

@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+from collections import defaultdict
 from types import SimpleNamespace
 
 import pytest
@@ -227,6 +228,47 @@ def test_talker2code2wav_full_payload_keeps_all_zero_codec_rows() -> None:
     assert "code_predictor_codes" not in payload
 
 
+def test_talker2code2wav_async_chunk_flushes_cached_tail_on_stop_token() -> None:
+    request_id = "codec_tail"
+    stop_token_id = 999
+    transfer_manager = SimpleNamespace(
+        code_prompt_token_ids=defaultdict(list),
+        put_req_chunk=defaultdict(int, {request_id: 2}),
+        connector=SimpleNamespace(
+            config={
+                "extra": {
+                    "initial_codec_chunk_frames": 4,
+                    "codec_chunk_frames": 25,
+                    "codec_left_context_frames": 25,
+                }
+            }
+        ),
+    )
+    transfer_manager.code_prompt_token_ids[request_id] = [
+        torch.tensor([[frame, frame + 100]], dtype=torch.long) for frame in range(50)
+    ]
+    request = SimpleNamespace(
+        external_req_id=request_id,
+        sampling_params=SimpleNamespace(
+            stop_token_ids=[stop_token_id],
+            stop_token_id=None,
+        ),
+    )
+
+    payload = q3.talker2code2wav_async_chunk(
+        transfer_manager,
+        {"codes": {"audio": torch.tensor([[stop_token_id, 0]], dtype=torch.long)}},
+        request,
+        is_finished=True,
+    )
+
+    assert payload is not None
+    assert payload.meta.finished.item() is True
+    # 4 initial frames and one 25-frame chunk were already emitted. The final
+    # payload contains 25 frames of left context plus the remaining 21 frames.
+    assert payload.codes.audio.numel() == (25 + 21) * 2
+
+
 def test_thinker2talker_full_payload_packs_complete_tensors() -> None:
     """Full-payload path drops the terminal thinker row before talker prefill."""
     request = SimpleNamespace(

tests/test_config_factory.py

@@ -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(
@@ -2162,6 +2164,43 @@ class TestBaseConfigInheritance:
 class TestPlatformOverrides:
     """Test platform-specific deploy config overrides."""
 
+    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)
+        # Talker KV is left to automatic sizing (no hard kv_cache_memory_bytes cap).
+        assert "kv_cache_memory_bytes" not in cuda_stages[1].yaml_engine_args
+
+        # 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
+        # also leave KV sizing automatic.
+        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 "kv_cache_memory_bytes" not in replica_stages[1].yaml_engine_args
+
+    def test_minicpmo_4_5_single_gpu_admits_four_sequences(self):
+        """Single-GPU shares leave room for Code2Wav peak activations at bs<=4."""
+        deploy = load_deploy_config(Path(get_deploy_config_path("minicpmo_4_5.yaml")))
+        assert [stage.max_num_seqs for stage in deploy.stages] == [4, 4, 4]
+        assert [stage.gpu_memory_utilization for stage in deploy.stages] == [
+            0.55,
+            0.15,
+            0.18,
+        ]
+
     def test_npu_overrides(self):
         deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
         if not deploy_path.exists():

tests/worker/test_native_duplex_hooks.py

@@ -276,16 +276,16 @@ def test_minicpmo_model_hook_same_speech_row_second_step_does_not_rearm_pending_
     next_logits[0, 7] = 30.0
     next_logits[0, 8] = -2.0
     next_logits[0, 10] = 20.0
+    original_next_logits = next_logits.clone()
 
     model.prepare_duplex_sampling(next_logits, SimpleNamespace(), (row,))
 
     assert state.pending_speech_context is False
     assert state.current_turn_ended is False
-    assert torch.isneginf(next_logits[0, 7])
-    assert next_logits[0, 8].item() == 30.0
+    assert torch.equal(next_logits, original_next_logits)
 
 
-def test_minicpmo_model_hook_mid_turn_speech_redirects_listen_to_tts_bos():
+def test_minicpmo_model_hook_mid_turn_speech_preserves_logits_for_post_sample_redirect():
     state = SimpleNamespace(
         current_turn_ended=False,
         last_terminator_token=None,
@@ -296,12 +296,11 @@ def test_minicpmo_model_hook_mid_turn_speech_redirects_listen_to_tts_bos():
     logits[0, 7] = 30.0
     logits[0, 8] = -2.0
     logits[0, 10] = 20.0
+    original_logits = logits.clone()
 
     model.prepare_duplex_sampling(logits, SimpleNamespace(), (row,))
 
-    assert torch.isneginf(logits[0, 7])
-    assert logits[0, 8].item() == 30.0
-    assert logits[0, 10].item() == 20.0
+    assert torch.equal(logits, original_logits)
 
 
 def test_minicpmo_model_hook_ignores_serving_new_user_turn_marker():
@@ -317,12 +316,12 @@ def test_minicpmo_model_hook_ignores_serving_new_user_turn_marker():
     logits = torch.zeros((1, 16), dtype=torch.float32)
     logits[0, 7] = 5.0
     logits[0, 10] = 20.0
+    original_logits = logits.clone()
     model.prepare_duplex_sampling(logits, SimpleNamespace(), (row,))
 
     assert state.current_turn_ended is False
     assert state.last_terminator_token is None
-    assert torch.isneginf(logits[0, 7])
-    assert logits[0, 8].item() == 5.0
+    assert torch.equal(logits, original_logits)
 
 
 def test_generic_ar_runner_builds_typed_duplex_sampling_rows():
@@ -613,6 +612,180 @@ def test_minicpmo_stage0_rejects_unknown_special_token_fallbacks():
         runtime._require_special_token_ids()
 
 
+def _minicpmo45_tokenizer_stub():
+    token_ids = {
+        token: index
+        for index, token in enumerate(
+            (
+                "<unit>",
+                "</unit>",
+                "<|listen|>",
+                "<|speak|>",
+                "<|tts_bos|>",
+                "<|tts_eos|>",
+                "<|tts_pad|>",
+                "<|chunk_eos|>",
+                "<|chunk_tts_eos|>",
+                "<|turn_eos|>",
+            ),
+            start=101,
+        )
+    }
+    return SimpleNamespace(
+        unk_token_id=0,
+        convert_tokens_to_ids=lambda token: token_ids.get(token, 0),
+        encode=lambda token, add_special_tokens=False: (
+            [token_ids[token]] if not add_special_tokens and token in token_ids else [0]
+        ),
+    )
+
+
+def _transformers_processor_stub(auto_processor):
+    class _AutoImageProcessor:
+        @staticmethod
+        def register(*_args, **_kwargs):
+            return None
+
+    return SimpleNamespace(
+        AutoImageProcessor=_AutoImageProcessor,
+        AutoProcessor=auto_processor,
+    )
+
+
+def test_minicpmo_stage0_loads_processor_from_hf_id(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    model_id = "openbmb/MiniCPM-o-4_5"
+    processor = SimpleNamespace(tokenizer=_minicpmo45_tokenizer_stub())
+    load_calls = []
+
+    class _AutoImageProcessor:
+        @staticmethod
+        def register(config_class, image_processor_class):
+            del image_processor_class
+            if isinstance(config_class, str):
+                raise AttributeError("'str' object has no attribute '__module__'")
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            load_calls.append((model_path, trust_remote_code))
+            _AutoImageProcessor.register("MiniCPMVImageProcessor", object)
+            return processor
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        SimpleNamespace(
+            AutoImageProcessor=_AutoImageProcessor,
+            AutoProcessor=_AutoProcessor,
+        ),
+    )
+
+    loaded = MiniCPMO45Stage0DuplexRuntime._load_processor_from_path(model_id)
+
+    assert loaded is processor
+    assert load_calls == [(model_id, True)]
+    with pytest.raises(AttributeError, match="__module__"):
+        _AutoImageProcessor.register("MiniCPMVImageProcessor", object)
+
+
+def test_minicpmo_stage0_processor_load_preserves_original_exception(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    original = ValueError("invalid processor config")
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            del cls, model_path, trust_remote_code
+            raise original
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        _transformers_processor_stub(_AutoProcessor),
+    )
+
+    with pytest.raises(RuntimeError, match="Failed to load MiniCPM-o duplex processor") as exc_info:
+        MiniCPMO45Stage0DuplexRuntime._load_processor_from_path("openbmb/MiniCPM-o-4_5")
+
+    assert exc_info.value.__cause__ is original
+
+
+def test_minicpmo_stage0_processor_requires_tokenizer(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            del cls, model_path, trust_remote_code
+            return object()
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        _transformers_processor_stub(_AutoProcessor),
+    )
+
+    with pytest.raises(RuntimeError, match="does not expose a tokenizer"):
+        MiniCPMO45Stage0DuplexRuntime._load_processor_from_path("openbmb/MiniCPM-o-4_5")
+
+
+def test_minicpmo_stage0_loaded_processor_validates_special_tokens(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    processor = SimpleNamespace(tokenizer=_minicpmo45_tokenizer_stub())
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            del cls, model_path, trust_remote_code
+            return processor
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        _transformers_processor_stub(_AutoProcessor),
+    )
+
+    runtime = MiniCPMO45Stage0DuplexRuntime(
+        SimpleNamespace(),
+        model_path="openbmb/MiniCPM-o-4_5",
+        device="cpu",
+    )
+
+    assert runtime.processor is processor
+    assert set(runtime._special_token_ids()) == {
+        "unit_token_id",
+        "unit_end_token_id",
+        "listen_token_id",
+        "speak_token_id",
+        "tts_bos_token_id",
+        "tts_eos_token_id",
+        "tts_pad_token_id",
+        "chunk_eos_token_id",
+        "chunk_tts_eos_token_id",
+        "turn_eos_token_id",
+    }
+
+
 def test_minicpmo_stage0_data_plane_prefill_matches_official_unit_format():
     import torch
 
@@ -1150,6 +1323,88 @@ def test_minicpmo_stage0_data_plane_final_first_chunk_does_not_add_silence_unit(
     )
 
 
+def test_minicpmo_stage0_final_does_not_promote_first_chunk_alignment_tail_to_unit():
+    import torch
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+        _MiniCPMO45Stage0SessionState,
+    )
+
+    class _StageModel(torch.nn.Module):
+        first_chunk_ms = 10
+        chunk_ms = 8
+        sample_rate = 1000
+
+        def __init__(self):
+            super().__init__()
+            self.embed = torch.nn.Embedding(256, 2)
+
+        def get_input_embeddings(self):
+            return self.embed
+
+        def get_audio_hidden_states(self, _data):
+            return [torch.tensor([[0.5, 0.5]], dtype=torch.float32)]
+
+    class _MelProcessor:
+        sample_rate = 1000
+        is_first = True
+
+        def get_config(self):
+            return {"effective_first_chunk_ms": 9}
+
+    class _Processor:
+        def __init__(self):
+            self._streaming_mel_processor = _MelProcessor()
+
+        def get_streaming_chunk_size(self):
+            return 9 if self._streaming_mel_processor.is_first else 8
+
+        def process_audio_streaming(self, audio, **_kwargs):
+            self._streaming_mel_processor.is_first = False
+            return {"audio_features": audio, "audio_feature_lens": [[len(audio)]]}
+
+    runtime = MiniCPMO45Stage0DuplexRuntime.__new__(MiniCPMO45Stage0DuplexRuntime)
+    runtime.stage_model = _StageModel()
+    runtime.thinker = runtime.stage_model
+    runtime.tokenizer = SimpleNamespace(
+        unk_token_id=0,
+        convert_tokens_to_ids=lambda token: {
+            "<unit>": 1,
+            "</unit>": 2,
+            "<|listen|>": 3,
+            "<|speak|>": 4,
+            "<|tts_bos|>": 5,
+            "<|tts_eos|>": 6,
+            "<|tts_pad|>": 7,
+            "<|chunk_eos|>": 8,
+            "<|chunk_tts_eos|>": 9,
+            "<|turn_eos|>": 10,
+            "<|audio|>": 11,
+        }.get(token, 0),
+        encode=lambda text, add_special_tokens=False: [],
+    )
+    runtime.processor = _Processor()
+    runtime.device = "cpu"
+    runtime._init_token_ids()
+    state = _MiniCPMO45Stage0SessionState(session_id="sid-first-chunk-alignment-tail")
+
+    first = runtime._stage_prefill_embeddings_only(state, np.zeros(8, dtype=np.float32), seq=1)
+    second = runtime._stage_prefill_embeddings_only(state, np.zeros(8, dtype=np.float32), seq=2)
+    final = runtime._stage_prefill_embeddings_only(
+        state,
+        np.zeros(8, dtype=np.float32),
+        seq=3,
+        final=True,
+    )
+
+    assert first["input_token_ids"].count(1) == 1
+    assert second["input_token_ids"].count(1) == 1
+    assert final["input_token_ids"].count(1) == 1
+    assert state.audio_chunk_idx == 3
+    assert len(state.audio_buffer) == 1
+
+
 def test_minicpmo_stage0_runtime_uses_loaded_vllm_embed_tokens_when_get_input_embeddings_is_broken():
     import torch
 
@@ -1319,7 +1574,7 @@ def test_minicpmo_stage0_native_sampler_penalizes_text_from_prior_chunk():
     model.thinker = SimpleNamespace(get_tokenizer=lambda: _Tokenizer())
     session_key = ("sid-cross-chunk-repetition", 0)
     state = _MiniCPMO45Stage0SessionState(session_id=session_key[0])
-    state.generated_text_tokens = [198] * 8
+    state.generated_tokens = [198] * 8
     model._minicpmo45_duplex_data_plane_helper = SimpleNamespace(sessions={session_key: state})
     model._minicpmo45_duplex_row_sessions = {0: session_key}
 
@@ -1381,7 +1636,7 @@ def test_minicpmo_stage0_native_sampler_matches_official_negative_logit_penalty(
     state = _MiniCPMO45Stage0SessionState(session_id=session_key[0])
     repeated = 198
     alternative = 1234
-    state.generated_text_tokens = [repeated]
+    state.generated_tokens = [repeated]
     model._minicpmo45_duplex_data_plane_helper = SimpleNamespace(sessions={session_key: state})
     model._minicpmo45_duplex_row_sessions = {0: session_key}
 
@@ -1405,7 +1660,7 @@ def test_minicpmo_stage0_native_sampler_matches_official_negative_logit_penalty(
     assert sampled.sampled_token_ids.tolist() == [[repeated]]
 
 
-def test_minicpmo_stage0_records_only_bounded_text_repetition_history():
+def test_minicpmo_stage0_records_bounded_model_policy_history():
     from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
         _MiniCPMO45Stage0SessionState,
     )
@@ -1432,15 +1687,15 @@ def test_minicpmo_stage0_records_only_bounded_text_repetition_history():
     }
 
     for sampled in range(1000, 1520):
-        model._record_minicpmo45_duplex_terminator(0, sampled, token_ids)
+        model._record_minicpmo45_duplex_generation_token(0, sampled)
 
     expected_history = list(range(1008, 1520))
-    assert state.generated_text_tokens == expected_history
+    assert state.generated_tokens == expected_history
 
     for sampled in token_ids.values():
-        model._record_minicpmo45_duplex_terminator(0, sampled, token_ids)
+        model._record_minicpmo45_duplex_generation_token(0, sampled)
 
-    assert state.generated_text_tokens == expected_history
+    assert state.generated_tokens == [*expected_history[len(token_ids) :], *token_ids.values()]
 
 
 def test_minicpmo_stage0_native_sampler_does_not_override_model_at_punctuation():
@@ -2211,3 +2466,165 @@ def test_minicpmo_stage0_session_context_includes_resolved_ref_audio():
 
     assert state.context_token_ids == [1, 2, 3, 151683, 151683, 4, 5]
     assert len(state.context_embeds) == 6
+
+
+# ---------------------------------------------------------------------------
+# Cross-platform wiring: every AR model runner must reach the duplex hook.
+#
+# ``GPUARModelRunner`` and ``NPUARModelRunner`` are siblings -- neither inherits
+# the other's ``_sample`` -- so the NPU runner once carried a copy of the GPU
+# ``_sample`` with the ``prepare_duplex_sampling`` call dropped.  That left the
+# MiniCPM-o 4.5 thinker without a row -> session map on Ascend, so every mid-turn
+# ``<|listen|>`` was emitted verbatim and the model stopped speaking after turn 1.
+#
+# These checks parse the source instead of importing it: the NPU runner needs
+# ``vllm_ascend``, which is absent on the CPU test images that run this file.
+# ---------------------------------------------------------------------------
+
+# method that must call it -> mixin method it must call
+_REQUIRED_DUPLEX_CALLS = {
+    "__init__": "_init_duplex_sampling_state",
+    "load_model": "_resolve_duplex_sampling_hook",
+    "_update_states": "_update_duplex_sampling_states",
+    "_sample": "_apply_duplex_sampling",
+}
+
+# The sampler ``_sample`` hands the prepared logits to.  ``_apply_duplex_sampling``
+# both masks those logits (force_listen) and publishes the row -> session map the
+# model's own sampler reads, so calling it *after* this is the same silent no-op as
+# not calling it at all.
+_MODEL_SAMPLER_CALL = "model_sample"
+
+
+def _repo_root():
+    from pathlib import Path
+
+    return Path(__file__).parents[2]
+
+
+def _ar_runner_sources():
+    """Every AR model runner in the repo, discovered rather than hand-listed.
+
+    A new platform runner is picked up automatically instead of silently escaping
+    the checks below.  ``vllm_omni/worker/gpu_ar_model_runner.py`` is named
+    separately because it does not live under ``platforms/``.
+    """
+    root = _repo_root()
+    paths = [root / "vllm_omni/worker/gpu_ar_model_runner.py"]
+    paths.extend(sorted(root.glob("vllm_omni/platforms/*/worker/*_ar_model_runner.py")))
+    return paths
+
+
+def _ar_runner_classdefs():
+    """``{class_name: (ClassDef, relative_path)}`` for every ``*ARModelRunner``."""
+    import ast
+
+    root = _repo_root()
+    found = {}
+    for path in _ar_runner_sources():
+        module = ast.parse(path.read_text())
+        for node in module.body:
+            if isinstance(node, ast.ClassDef) and node.name.endswith("ARModelRunner"):
+                found[node.name] = (node, str(path.relative_to(root)))
+    assert found, "no AR model runners discovered; the glob in _ar_runner_sources() is stale"
+    return found
+
+
+def _method_calls(method):
+    """``{attribute_name: first lineno}`` for every ``x.y(...)`` call in ``method``."""
+    import ast
+
+    calls = {}
+    for node in ast.walk(method):
+        if isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute):
+            calls.setdefault(node.func.attr, node.lineno)
+    return calls
+
+
+def _wires_duplex_sampling_directly(classdef) -> bool:
+    import ast
+
+    bases = {base.id for base in classdef.bases if isinstance(base, ast.Name)}
+    if "DuplexSamplingRunnerMixin" not in bases:
+        return False
+    methods = {node.name: node for node in classdef.body if isinstance(node, ast.FunctionDef)}
+    return all(
+        method_name in methods and required_call in _method_calls(methods[method_name])
+        for method_name, required_call in _REQUIRED_DUPLEX_CALLS.items()
+    )
+
+
+def test_every_ar_runner_reaches_the_duplex_sampling_hook():
+    """Each AR runner must wire the hook itself or inherit from one that does.
+
+    ``XPUARModelRunner`` derives from ``GPUARModelRunner`` and so is wired for free;
+    ``NPUARModelRunner`` derives from ``OmniNPUModelRunner``, which is not an AR
+    runner, so it has to do the wiring itself.  Asserting every discovered class
+    wires directly would wrongly fail the XPU runner.
+    """
+    import ast
+
+    classdefs = _ar_runner_classdefs()
+    wired = {name for name, (classdef, _path) in classdefs.items() if _wires_duplex_sampling_directly(classdef)}
+
+    for class_name, (classdef, path) in sorted(classdefs.items()):
+        if class_name in wired:
+            continue
+        base_names = {base.id for base in classdef.bases if isinstance(base, ast.Name)}
+        assert base_names & wired, (
+            f"{path}::{class_name} neither wires the duplex sampling hook nor derives from a "
+            f"runner that does (wired runners: {sorted(wired)}).  Inherit DuplexSamplingRunnerMixin "
+            f"and call {sorted(_REQUIRED_DUPLEX_CALLS.values())} from "
+            f"{sorted(_REQUIRED_DUPLEX_CALLS)} respectively."
+        )
+
+
+@pytest.mark.parametrize("class_name", ["GPUARModelRunner", "NPUARModelRunner"])
+def test_ar_runner_calls_every_duplex_sampling_hook_site(class_name: str):
+    import ast
+
+    classdefs = _ar_runner_classdefs()
+    assert class_name in classdefs, f"{class_name} not found by _ar_runner_sources()"
+    classdef, path = classdefs[class_name]
+
+    bases = {base.id for base in classdef.bases if isinstance(base, ast.Name)}
+    assert "DuplexSamplingRunnerMixin" in bases, (
+        f"{path}::{class_name} must inherit DuplexSamplingRunnerMixin so the duplex "
+        f"sampling hook is wired the same way on every platform"
+    )
+
+    methods = {node.name: node for node in classdef.body if isinstance(node, ast.FunctionDef)}
+    for method_name, required_call in _REQUIRED_DUPLEX_CALLS.items():
+        method = methods.get(method_name)
+        assert method is not None, f"{path}::{class_name}.{method_name} is missing; it must call {required_call}"
+        calls = _method_calls(method)
+        assert required_call in calls, f"{path}::{class_name}.{method_name} must call self.{required_call}"
+
+
+@pytest.mark.parametrize("class_name", ["GPUARModelRunner", "NPUARModelRunner"])
+def test_ar_runner_applies_duplex_sampling_before_the_model_sampler(class_name: str):
+    """Ordering is load-bearing: the hook must run before the sampler reads the logits."""
+    import ast
+
+    classdef, path = _ar_runner_classdefs()[class_name]
+    sample = next(node for node in classdef.body if isinstance(node, ast.FunctionDef) and node.name == "_sample")
+
+    apply_lineno = _method_calls(sample).get("_apply_duplex_sampling")
+    sampler_lineno = next(
+        (
+            node.lineno
+            for node in ast.walk(sample)
+            if isinstance(node, ast.Call) and isinstance(node.func, ast.Name) and node.func.id == _MODEL_SAMPLER_CALL
+        ),
+        None,
+    )
+    assert apply_lineno is not None, f"{path}::{class_name}._sample must call self._apply_duplex_sampling"
+    assert sampler_lineno is not None, (
+        f"{path}::{class_name}._sample no longer calls {_MODEL_SAMPLER_CALL}(); update _MODEL_SAMPLER_CALL "
+        f"so the ordering check keeps testing something"
+    )
+    assert apply_lineno < sampler_lineno, (
+        f"{path}::{class_name}._sample calls self._apply_duplex_sampling at line {apply_lineno}, after "
+        f"{_MODEL_SAMPLER_CALL}() at line {sampler_lineno}.  The hook masks the logits and publishes the "
+        f"row -> session map the sampler reads, so running it afterwards is a silent no-op."
+    )

vllm_omni/benchmarks/data_modules/daily_omni_dataset.py

@@ -7,7 +7,8 @@ Dataset source: https://huggingface.co/datasets/liarliar/Daily-Omni
 
 Supports loading QA metadata from:
 - Local JSON file (``qa_json_path``): recommended for offline/air-gapped environments
-- HuggingFace datasets (``dataset_path``): legacy online mode
+- HuggingFace Hub id (``dataset_path``): resolved with ``snapshot_download`` to the repo's
+  ``qa.json``, which reads the hub cache and therefore also works under ``HF_HUB_OFFLINE``
 
 Video/audio files normally come from extracted ``Videos.tar``. When ``--daily-omni-video-dir``
 is not set, the first request that needs on-disk media downloads that archive from the Hugging Face
@@ -24,9 +25,9 @@ Why ``BenchmarkDataset`` instead of ``HuggingFaceDataset``?
     This class therefore inherits only ``BenchmarkDataset`` (minimal: ``dataset_path``,
     ``random_seed``, ``self.data``) and implements **two explicit loaders**:
     ``_load_from_local_json`` (default path for air-gapped runs) and ``_load_from_huggingface``
-    (optional legacy path for users who prefer ``datasets`` + Hub cache). The latter is **not**
+    (Hub ids, resolved to the repo's ``qa.json`` through the hub cache). The latter is **not**
     inheritance; it is the same Hub rows as before, factored into a helper so one class can
-    serve both deployment modes without mandatory ``datasets`` when using ``qa_json_path``.
+    serve both deployment modes without mandatory ``datasets``.
 
 Usage:
     from vllm_omni.benchmarks.data_modules.daily_omni_dataset import DailyOmniDataset
@@ -38,7 +39,7 @@ Usage:
         random_seed=42,
     )
 
-    # HuggingFace mode (legacy, requires network)
+    # Hub id mode (works offline once the repo is in the HF hub cache)
     dataset = DailyOmniDataset(
         dataset_path="liarliar/Daily-Omni",
         dataset_split="train",
@@ -52,11 +53,14 @@ Usage:
 """
 
 import base64
+import io
 import json
 import logging
+import math
 import os
 import shutil
 import tarfile
+import wave
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Literal
@@ -77,6 +81,10 @@ except ImportError:
 
 logger = logging.getLogger(__name__)
 
+# Official MiniCPM ``get_video_frame_audio_segments`` default (env ``MAX_NUM_FRAMES``).
+_MINICPM_MAX_NUM_FRAMES = int(os.environ.get("MAX_NUM_FRAMES", "64"))
+_MINICPM_AUDIO_SR = 16000
+
 
 def _daily_omni_hf_cache_root() -> Path:
     return Path(os.environ.get("HF_HOME", Path.home() / ".cache" / "huggingface")).expanduser().resolve()
@@ -103,47 +111,17 @@ def _daily_omni_find_videos_root_in_extract(tmp: Path) -> Path:
     )
 
 
-def ensure_daily_omni_hub_videos_dir(repo_id: str) -> Path:
-    """Download ``Videos.tar`` from the Hugging Face dataset repo and return the ``Videos`` root.
+def _extract_daily_omni_videos_tar(tar_path: Path, cache_key: str) -> Path:
+    """Extract ``Videos.tar`` once into the shared media cache and return the ``Videos`` root.
 
-    The returned path matches ``--daily-omni-video-dir`` (directory containing ``{{video_id}}/``).
-
-    Cached under ``HF_HOME`` / ``vllm_omni/daily_omni_media/<repo>``. Reuses extraction when the
-    tarball fingerprint matches.
-
-    Raises:
-        ImportError: if ``huggingface_hub`` is not installed.
-        FileNotFoundError / RuntimeError: if the archive is missing or malformed.
+    Cached under ``HF_HOME`` / ``vllm_omni/daily_omni_media/<cache_key>``. Reuses a previous
+    extraction when the tarball fingerprint matches.
     """
-    rid = (repo_id or "").strip()
-    if not rid:
-        raise ValueError("repo_id is required to download Daily-Omni Videos.tar")
-
-    try:
-        from huggingface_hub import hf_hub_download
-    except ImportError as e:
-        raise ImportError(
-            "Daily-Omni Hub media download requires huggingface_hub. "
-            "Install it (e.g. with vLLM) or provide --daily-omni-video-dir with a local extract."
-        ) from e
-
-    safe = rid.replace("/", "__").replace("\\", "_")
+    safe = cache_key.replace("/", "__").replace("\\", "_")
     staging_root = _daily_omni_hf_cache_root() / "vllm_omni" / "daily_omni_media" / safe
     videos_dir = staging_root / "Videos"
     marker = staging_root / ".videos_extracted"
 
-    tar_path: Path | None = None
-    for fname in ("Videos.tar", "videos.tar"):
-        try:
-            tar_path = Path(hf_hub_download(repo_id=rid, filename=fname, repo_type="dataset"))
-            break
-        except Exception:
-            continue
-    if tar_path is None or not tar_path.is_file():
-        raise FileNotFoundError(
-            f"Could not download Videos.tar from Hugging Face dataset {rid!r} (tried Videos.tar / videos.tar)."
-        )
-
     fp = _daily_omni_tar_fingerprint(tar_path)
     if marker.is_file() and videos_dir.is_dir():
         try:
@@ -160,7 +138,7 @@ def ensure_daily_omni_hub_videos_dir(repo_id: str) -> Path:
     shutil.rmtree(work, ignore_errors=True)
     work.mkdir(parents=True)
     try:
-        logger.info("Extracting Daily-Omni Videos.tar from %s (repo=%s)", tar_path, rid)
+        logger.info("Extracting Daily-Omni Videos.tar from %s (cache_key=%s)", tar_path, cache_key)
         with tarfile.open(tar_path, "r:*") as tf:
             tf.extractall(path=work, filter="data")
         found = _daily_omni_find_videos_root_in_extract(work)
@@ -171,10 +149,168 @@ def ensure_daily_omni_hub_videos_dir(repo_id: str) -> Path:
         shutil.rmtree(work, ignore_errors=True)
 
     marker.write_text(fp, encoding="utf-8")
-    logger.info("Daily-Omni Hub media ready at %s", videos_dir)
+    logger.info("Daily-Omni media ready at %s", videos_dir)
     return videos_dir
 
 
+def ensure_daily_omni_hub_videos_dir(repo_id: str) -> Path:
+    """Download ``Videos.tar`` from the Hugging Face dataset repo and return the ``Videos`` root.
+
+    The returned path matches ``--daily-omni-video-dir`` (directory containing ``{{video_id}}/``).
+
+    Raises:
+        ImportError: if ``huggingface_hub`` is not installed.
+        FileNotFoundError / RuntimeError: if the archive is missing or malformed.
+    """
+    rid = (repo_id or "").strip()
+    if not rid:
+        raise ValueError("repo_id is required to download Daily-Omni Videos.tar")
+
+    try:
+        from huggingface_hub import hf_hub_download
+    except ImportError as e:
+        raise ImportError(
+            "Daily-Omni Hub media download requires huggingface_hub. "
+            "Install it (e.g. with vLLM) or provide --daily-omni-video-dir with a local extract."
+        ) from e
+
+    tar_path: Path | None = None
+    for fname in ("Videos.tar", "videos.tar"):
+        try:
+            tar_path = Path(hf_hub_download(repo_id=rid, filename=fname, repo_type="dataset"))
+            break
+        except Exception:
+            continue
+    if tar_path is None or not tar_path.is_file():
+        raise FileNotFoundError(
+            f"Could not download Videos.tar from Hugging Face dataset {rid!r} (tried Videos.tar / videos.tar)."
+        )
+
+    return _extract_daily_omni_videos_tar(tar_path, rid)
+
+
+def _resolve_hf_cache_snapshot(path: Path) -> Path:
+    """Map a HF hub cache dir (``datasets--org--name``) to its checked-out snapshot dir.
+
+    Any other directory is returned unchanged.
+    """
+    snapshots = path / "snapshots"
+    if not snapshots.is_dir():
+        return path
+    ref = path / "refs" / "main"
+    if ref.is_file():
+        try:
+            target = snapshots / ref.read_text(encoding="utf-8").strip()
+        except OSError:
+            target = snapshots
+        if target.is_dir():
+            return target
+    revisions = [p for p in snapshots.iterdir() if p.is_dir()]
+    if revisions:
+        return max(revisions, key=lambda p: p.stat().st_mtime)
+    return path
+
+
+#: QA metadata pulled by :func:`ensure_daily_omni_hub_root`. ``Videos.tar`` is deliberately left
+#: out: it is ~60GB and only :func:`ensure_daily_omni_hub_videos_dir` needs it, lazily, once a
+#: request actually asks for media.
+_DAILY_OMNI_QA_ALLOW_PATTERNS = ["qa.json", "QA.json"]
+
+
+def ensure_daily_omni_hub_root(repo_id: str) -> Path:
+    """Resolve a Daily-Omni Hub id to its snapshot directory, downloading QA metadata if needed.
+
+    Mirrors :func:`...videomme_dataset.ensure_videomme_hub_root` and Seed-TTS: ``snapshot_download``
+    resolves the repo through the **hub cache**, so an air-gapped run (``HF_HUB_OFFLINE=1``) keeps
+    working as long as the dataset was pre-staged there. ``datasets.load_dataset`` cannot do this —
+    it only consults the separate *datasets* cache, which a ``hf download`` never populates.
+
+    Raises:
+        ImportError: if ``huggingface_hub`` is not installed.
+        ValueError: if ``repo_id`` is empty.
+    """
+    rid = (repo_id or "").strip()
+    if not rid:
+        raise ValueError("repo_id is required to download Daily-Omni from Hugging Face")
+
+    try:
+        from huggingface_hub import snapshot_download
+    except ImportError as e:
+        raise ImportError(
+            "Install huggingface_hub to load Daily-Omni from the Hub, or pass a local "
+            "--daily-omni-qa-json / --daily-omni-video-dir mirror."
+        ) from e
+
+    cache = snapshot_download(
+        repo_id=rid,
+        repo_type="dataset",
+        allow_patterns=_DAILY_OMNI_QA_ALLOW_PATTERNS,
+    )
+    root = _resolve_hf_cache_snapshot(Path(cache).resolve())
+    logger.info("Daily-Omni Hub snapshot ready at %s (repo=%s)", root, rid)
+    return root
+
+
+def _daily_omni_qa_load_error(
+    repo_id: str | None,
+    snapshot_error: Exception | None,
+    datasets_error: Exception,
+) -> RuntimeError:
+    """Compose an actionable error after both QA-metadata sources failed.
+
+    Offline CI hits this when the dataset is missing from the hub cache *or* staged under a
+    different repo id, so the message names both attempts and the flags that bypass them.
+    """
+    parts = [f"Could not load Daily-Omni QA metadata for {repo_id!r}."]
+    if snapshot_error is not None:
+        parts.append(f"huggingface_hub snapshot_download failed: {type(snapshot_error).__name__}: {snapshot_error}")
+    parts.append(f"datasets.load_dataset failed: {type(datasets_error).__name__}: {datasets_error}")
+    parts.append(
+        "For offline runs, pre-stage the dataset in the HF hub cache "
+        "(`hf download --repo-type dataset <repo>`) and make sure the repo id matches, "
+        "or pass --daily-omni-qa-json / --daily-omni-video-dir pointing at a local mirror."
+    )
+    return RuntimeError(" ".join(parts))
+
+
+def resolve_daily_omni_local_root(dataset_path: str | None) -> Path | None:
+    """Return the local Daily-Omni mirror root for ``dataset_path``, or ``None`` if not local.
+
+    Accepts a plain directory (``qa.json`` + ``Videos/`` or ``Videos.tar``, i.e. the layout of the
+    Hub repo) as well as a HuggingFace hub cache directory (``datasets--liarliar--Daily-Omni``),
+    which is resolved to its snapshot. Hub ids such as ``liarliar/Daily-Omni`` return ``None``.
+    """
+    raw = (dataset_path or "").strip()
+    if not raw:
+        return None
+    path = Path(raw).expanduser()
+    if not path.is_dir():
+        return None
+    return _resolve_hf_cache_snapshot(path.resolve())
+
+
+def daily_omni_local_qa_json(root: Path) -> Path | None:
+    """Locate ``qa.json`` inside a local Daily-Omni mirror (``None`` when absent)."""
+    for name in ("qa.json", "QA.json"):
+        candidate = root / name
+        if candidate.is_file():
+            return candidate
+    return None
+
+
+def daily_omni_local_videos_dir(root: Path) -> Path | None:
+    """Return the ``Videos`` root of a local mirror, extracting ``Videos.tar`` when needed."""
+    for name in ("Videos", "videos"):
+        candidate = root / name
+        if candidate.is_dir():
+            return candidate
+    for name in ("Videos.tar", "videos.tar"):
+        tar_path = root / name
+        if tar_path.is_file():
+            return _extract_daily_omni_videos_tar(tar_path, f"local__{root.name}")
+    return None
+
+
 class _ListDatasetIterator:
     """Simple iterator wrapper around a list to mimic HuggingFace streaming dataset behavior."""
 
@@ -203,12 +339,130 @@ class _ListDatasetIterator:
 # Aligns with Lliar-liar/Daily-Omni CLI ``--input_mode`` (test_model/*/testmodel.py).
 DailyOmniInputMode = Literal["all", "visual", "audio"]
 
+# How multimodal parts are packed into OpenAI chat messages:
+# - ``qwen``: one ``video_url`` + one ``audio_url`` (upstream Daily-Omni / Qwen protocol).
+# - ``minicpm-interleave``: 1fps ``[image_i, audio_i, ...]`` matching OpenBMB MiniCPM-o
+#   ``get_video_frame_audio_segments`` + interleaved omni contents (paper ~80% recipe).
+DailyOmniPackMode = Literal["qwen", "minicpm-interleave"]
+
 # ``build_conversation()`` in Daily-Omni ``test_model/Qwen2.5-Omni/testmodel.py`` (verbatim).
 DAILY_OMNI_SYSTEM_TEXT = (
     "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, "
     "capable of perceiving auditory and visual inputs, as well as generating text and speech."
 )
 
+# MiniCPM-o ``get_sys_prompt(mode="omni")`` without ref-audio is an empty system string.
+MINICPM_OMNI_SYSTEM_TEXT = ""
+
+
+def _uniform_sample_indices(n: int, k: int) -> list[int]:
+    """Evenly pick ``k`` indices from ``0..n-1`` (MiniCPM ``uniform_sample`` on index lists)."""
+    if k <= 0:
+        return []
+    if n <= k:
+        return list(range(n))
+    # Match numpy linspace(..., dtype=int) used upstream.
+    try:
+        import numpy as np
+
+        return [int(i) for i in np.linspace(0, n - 1, k, dtype=int)]
+    except ImportError:
+        if k == 1:
+            return [0]
+        return [int(round(i * (n - 1) / (k - 1))) for i in range(k)]
+
+
+def _probe_video_frames_and_fps(video_path: Path) -> tuple[int, float]:
+    """Frame count and average FPS, replacing ``len(vr)`` / ``vr.get_avg_fps()``.
+
+    PyAV is used instead of ``decord`` because ``decord`` ships x86_64-only Linux wheels
+    and is unmaintained, while ``av`` is already a first-class requirement and has
+    aarch64 wheels (Ascend hosts).
+    """
+    import av
+
+    with av.open(str(video_path)) as container:
+        stream = container.streams.video[0]
+        avg_fps = float(stream.average_rate) if stream.average_rate else 0.0
+        num_frames = stream.frames or 0
+        if num_frames <= 0:
+            # Containers such as raw webm leave ``frames`` at 0; derive it from duration.
+            duration: float | None = None
+            if stream.duration is not None and stream.time_base is not None:
+                duration = float(stream.duration * stream.time_base)
+            elif container.duration is not None:
+                duration = container.duration / av.time_base
+            if duration and avg_fps:
+                num_frames = int(math.floor(duration * avg_fps))
+        if num_frames <= 0:
+            # Last resort for streams with no usable metadata at all.
+            stream.thread_type = "AUTO"
+            num_frames = sum(1 for _ in container.decode(stream))
+    if num_frames <= 0:
+        raise ValueError(f"No decodable video frames in {video_path}")
+    return num_frames, avg_fps or 1.0
+
+
+def _decode_video_frames_rgb(video_path: Path, frame_idx: list[int]) -> list[Any]:
+    """Decode presentation-order ``frame_idx`` into RGB ``PIL.Image``s.
+
+    Mirrors ``decord.VideoReader.get_batch(frame_idx)``: one image per requested index,
+    in the requested order, duplicates repeated. Decoding is a single forward pass
+    (no seeking) so the mapping from index to picture matches decord's indexing.
+    """
+    import av
+
+    wanted = sorted({int(i) for i in frame_idx})
+    if not wanted:
+        return []
+
+    decoded: dict[int, Any] = {}
+    with av.open(str(video_path)) as container:
+        stream = container.streams.video[0]
+        stream.thread_type = "AUTO"
+        cursor = 0
+        for position, frame in enumerate(container.decode(stream)):
+            if wanted[cursor] > position:
+                continue
+            image = frame.to_image()
+            # ``<=`` also absorbs indices that overshoot the real frame count.
+            while cursor < len(wanted) and wanted[cursor] <= position:
+                decoded[wanted[cursor]] = image
+                cursor += 1
+            if cursor >= len(wanted):
+                break
+
+    if not decoded:
+        raise ValueError(f"Decoded no frames from {video_path} for indices {wanted[:8]}")
+    # Metadata may promise more frames than the stream actually decodes; clamp to the last one.
+    last_image = decoded[max(decoded)]
+    return [decoded.get(int(i), last_image) for i in frame_idx]
+
+
+def _numpy_to_wav_bytes(audio_np: Any, sample_rate: int = _MINICPM_AUDIO_SR) -> bytes:
+    """Encode a mono float/int numpy waveform as 16-bit PCM WAV bytes."""
+    import numpy as np
+
+    arr = np.asarray(audio_np)
+    if arr.ndim > 1:
+        arr = arr.reshape(-1)
+    if arr.dtype != np.int16:
+        arr = np.clip(arr.astype(np.float64), -1.0, 1.0)
+        arr = (arr * 32767.0).astype(np.int16)
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as wf:
+        wf.setnchannels(1)
+        wf.setsampwidth(2)
+        wf.setframerate(sample_rate)
+        wf.writeframes(arr.tobytes())
+    return buf.getvalue()
+
+
+def _pil_to_jpeg_bytes(image: Any, quality: int = 85) -> bytes:
+    buf = io.BytesIO()
+    image.save(buf, format="JPEG", quality=quality)
+    return buf.getvalue()
+
 
 @dataclass
 class DailyOmniSampleRequest(SampleRequest):
@@ -242,7 +496,8 @@ class DailyOmniDataset(BenchmarkDataset):
 
     QA metadata can be loaded from:
     - Local JSON file (``qa_json_path``): recommended for offline/air-gapped environments
-    - HuggingFace datasets (``dataset_path``): legacy online mode
+    - HuggingFace Hub id (``dataset_path``): the repo's ``qa.json`` via ``snapshot_download``
+      (hub-cache backed, offline-safe); ``datasets.load_dataset`` only as a parquet fallback
 
     Video/audio files normally come from extracted ``Videos.tar``. When ``video_dir`` is not set,
     the first sample that needs on-disk media downloads that archive from the Hugging Face dataset
@@ -253,7 +508,8 @@ class DailyOmniDataset(BenchmarkDataset):
         qa_json_path: Path to local qa.json file (offline mode, preferred). When provided,
             ``dataset_path`` and ``dataset_split`` are ignored.
         dataset_path: HuggingFace dataset path (e.g., "liarliar/Daily-Omni"). Used only if
-            ``qa_json_path`` is not provided (legacy online mode).
+            ``qa_json_path`` is not provided; resolved through the hub cache, so a pre-staged
+            repo also loads with ``HF_HUB_OFFLINE=1``.
         dataset_split: Dataset split to use (default: "train"). Used only in online mode.
         random_seed: Random seed for shuffling
         video_dir: Directory containing extracted video files (default: None; may be filled lazily
@@ -262,6 +518,11 @@ class DailyOmniDataset(BenchmarkDataset):
             ``all`` — video + WAV (default; official audio-visual protocol);
             ``visual`` — video only;
             ``audio`` — extracted WAV only (requires ``{video_id}/{video_id}_audio.wav`` under ``video_dir``).
+        pack_mode: Multimodal packing recipe:
+            ``qwen`` — one ``video_url`` + optional ``audio_url`` (default; Daily-Omni/Qwen);
+            ``minicpm-interleave`` — MiniCPM-o paper recipe: 1fps frames interleaved with
+            matching 1s audio segments as ``image_url``/``audio_url`` pairs (required to approach
+            OpenBMB ~80% Daily-Omni; needs local video+WAV under ``video_dir``).
         max_duration_seconds: Reserved for future ffprobe-based filtering; currently **not applied**
             when building requests (metadata ``video_duration`` is still passed through for eval).
         dataset_subset: Optional HuggingFace subset name (``load_dataset(..., name=...)``); used by bench
@@ -270,7 +531,8 @@ class DailyOmniDataset(BenchmarkDataset):
         inline_local_video: If True, embed local MP4 as ``data:video/mp4;base64,...`` in requests so
             the API server does not need ``--allowed-local-media-path`` (large JSON; use for small runs).
             When ``input_mode`` is ``audio`` or ``all``, local WAV is embedded the same way
-            (``data:audio/wav;base64,...``).
+            (``data:audio/wav;base64,...``). For ``minicpm-interleave``, embeds JPEG/WAV segment
+            data URLs instead of the raw MP4.
         trust_remote_code: Whether to trust remote code when loading HuggingFace dataset
             (online mode only).
     """
@@ -291,6 +553,7 @@ class DailyOmniDataset(BenchmarkDataset):
         random_seed: int = 0,
         video_dir: str | None = None,
         input_mode: DailyOmniInputMode = "all",
+        pack_mode: DailyOmniPackMode = "qwen",
         inline_local_video: bool = False,
         trust_remote_code: bool = False,
         max_duration_seconds: float | None = None,
@@ -300,6 +563,8 @@ class DailyOmniDataset(BenchmarkDataset):
     ) -> None:
         if input_mode not in ("all", "visual", "audio"):
             raise ValueError(f"input_mode must be 'all', 'visual', or 'audio', got {input_mode!r}")
+        if pack_mode not in ("qwen", "minicpm-interleave"):
+            raise ValueError(f"pack_mode must be 'qwen' or 'minicpm-interleave', got {pack_mode!r}")
 
         # Validate arguments: need either local JSON or HF path
         if qa_json_path is None and dataset_path is None:
@@ -318,11 +583,14 @@ class DailyOmniDataset(BenchmarkDataset):
         self.video_dir = Path(video_dir) if video_dir else None
         self.inline_local_video = inline_local_video
         self.input_mode: DailyOmniInputMode = input_mode
+        self.pack_mode: DailyOmniPackMode = pack_mode
         self.max_duration_seconds = max_duration_seconds
         self.trust_remote_code = trust_remote_code
 
         #: In-process cache of ffprobe durations only (no disk persistence).
         self._video_durations: dict[str, float] = {}
+        #: Cache of MiniCPM 1fps interleaved expansions: video_id -> content parts.
+        self._minicpm_interleave_cache: dict[str, list[dict[str, Any]]] = {}
 
         # Initialize parent BenchmarkDataset
         super().__init__(
@@ -336,11 +604,13 @@ class DailyOmniDataset(BenchmarkDataset):
 
         # Verify dataset info
         logger.info(
-            "Loaded Daily-Omni dataset: mode=%s, source=%s, random_seed=%d, input_mode=%s, max_duration=%s",
+            "Loaded Daily-Omni dataset: mode=%s, source=%s, random_seed=%d, "
+            "input_mode=%s, pack_mode=%s, max_duration=%s",
             "local_json" if self.qa_json_path else "huggingface",
             str(self.qa_json_path) if self.qa_json_path else f"{dataset_path}/{dataset_split}",
             random_seed,
             input_mode,
+            pack_mode,
             f"{max_duration_seconds}s" if max_duration_seconds else "unlimited",
         )
 
@@ -397,21 +667,42 @@ class DailyOmniDataset(BenchmarkDataset):
         self.data = _ListDatasetIterator(data)
 
     def _load_from_huggingface(self) -> None:
-        """Load QA rows via ``datasets.load_dataset`` (legacy / convenience path).
+        """Load QA rows for a Hub id, preferring the repo's ``qa.json`` over ``datasets``.
 
-        Kept for backward compatibility: callers can still pass ``dataset_path=liarliar/Daily-Omni``
-        and get the same parquet-backed rows as the Hub dataset card, with streaming (or
-        non-streaming if ``no_stream=True``) and shuffle.
+        The Hub repo ships ``qa.json`` (1,197 rows) next to ``Videos.tar``, so the primary path is
+        :func:`ensure_daily_omni_hub_root` → :meth:`_load_from_local_json`: ``snapshot_download``
+        reads the hub cache and therefore works under ``HF_HUB_OFFLINE``, the same contract as
+        Seed-TTS and Video-MME. ``datasets.load_dataset`` remains only as a fallback for mirrors
+        that publish parquet rows instead of ``qa.json``; it consults the *datasets* cache rather
+        than the hub cache and streams shards over HTTP, so it cannot serve air-gapped runs.
 
         This is intentionally **not** implemented by subclassing ``HuggingFaceDataset``: that base
         always runs Hub ``load_dataset`` from its constructor and expects a Hub id as the primary
         API; Daily-Omni instead chooses the source in ``load_data()`` (JSON vs Hub) while sharing
         one ``sample()`` / request-building implementation for both.
         """
+        snapshot_error: Exception | None = None
+        try:
+            root = ensure_daily_omni_hub_root(self.dataset_path)
+        except Exception as e:  # noqa: BLE001 - any failure just falls through to `datasets`
+            snapshot_error = e
+        else:
+            qa_json = daily_omni_local_qa_json(root)
+            if qa_json is not None:
+                self.qa_json_path = qa_json
+                logger.info("Loading Daily-Omni QA metadata from Hub snapshot: %s", qa_json)
+                self._load_from_local_json()
+                return
+            logger.info(
+                "Hub snapshot %s ships no qa.json; falling back to `datasets.load_dataset`",
+                root,
+            )
+
         if load_dataset is None:
-            raise ImportError(
-                "datasets library is required for HuggingFace mode. "
-                "Install with: pip install datasets, or use local JSON mode instead."
+            raise _daily_omni_qa_load_error(
+                self.dataset_path,
+                snapshot_error,
+                ImportError("datasets library is not installed (pip install datasets)"),
             )
 
         load_kw: dict[str, Any] = {
@@ -421,7 +712,10 @@ class DailyOmniDataset(BenchmarkDataset):
         }
         if self.dataset_subset is not None:
             load_kw["name"] = self.dataset_subset
-        ds = load_dataset(self.dataset_path, **load_kw)
+        try:
+            ds = load_dataset(self.dataset_path, **load_kw)
+        except Exception as e:
+            raise _daily_omni_qa_load_error(self.dataset_path, snapshot_error, e) from e
         if not getattr(self, "disable_shuffle", False):
             ds = ds.shuffle(seed=self.random_seed)
         self.data = ds
@@ -575,7 +869,8 @@ class DailyOmniDataset(BenchmarkDataset):
         messages = self._build_daily_omni_openai_messages(mm_payload, question, choice)
         user_text = self._official_daily_omni_user_prompt(question, choice)
         # Text-only length estimate (same as before: no MM token count in bench).
-        prompt_len = len(tokenizer.encode(f"{DAILY_OMNI_SYSTEM_TEXT}\n{user_text}"))
+        system_text = self._system_text_for_pack_mode()
+        prompt_len = len(tokenizer.encode(f"{system_text}\n{user_text}" if system_text else user_text))
 
         return DailyOmniSampleRequest(
             prompt=user_text,
@@ -766,10 +1061,11 @@ class DailyOmniDataset(BenchmarkDataset):
         video_id: str,
         video_url: str | None,
     ) -> tuple[dict[str, Any] | list[dict[str, Any]] | None, dict[str, Any] | None, Literal["first", "last"]]:
-        """Build ``multi_modal_data`` and request extras for the active ``input_mode``.
+        """Build multimodal OpenAI content parts and request extras for the active modes."""
+        if self.pack_mode == "minicpm-interleave":
+            return self._compose_minicpm_interleaved_multimodal(video_id, video_url)
 
-        Mirrors upstream Daily-Omni: separate video + WAV with ``use_audio_in_video=False``.
-        """
+        # Qwen / upstream Daily-Omni: separate video + WAV with ``use_audio_in_video=False``.
         extra: dict[str, Any] = {"mm_processor_kwargs": {"use_audio_in_video": False}}
         mode = self.input_mode
 
@@ -787,6 +1083,215 @@ class DailyOmniDataset(BenchmarkDataset):
             return None, None, "first"
         return [v, a], extra, "first"
 
+    def _compose_minicpm_interleaved_multimodal(
+        self,
+        video_id: str,
+        video_url: str | None,
+    ) -> tuple[list[dict[str, Any]] | None, dict[str, Any] | None, Literal["first", "last"]]:
+        """Pack media like MiniCPM-o official omni chat: 1fps ``[frame_i, audio_i, ...]``.
+
+        OpenBMB's HF / utils path expands ``video_url`` with ``use_audio=True`` into interleaved
+        PIL frames and 1s audio ndarrays. vLLM-Omni does not implement that expansion, so the
+        bench client builds ordered ``image_url`` / ``audio_url`` parts instead. MiniCPM-o 5.0's
+        OpenAI chat template preserves content-part order, so placeholders stay time-aligned.
+
+        ``max_slice_nums=1`` and ``use_image_id=False`` mirror MiniCPM-o's omni path
+        (``modeling_minicpmo.py`` passes both for interleaved frames): slicing every frame with
+        the config default of 9 would blow past ``max_model_len``, and ``<image_id>N</image_id>``
+        prefixes would label each frame as a separate picture instead of a video timeline.
+        """
+        del video_url  # Hub HTTP URLs are not used; local extract is required.
+        extra: dict[str, Any] = {
+            "mm_processor_kwargs": {
+                "use_audio_in_video": False,
+                "max_slice_nums": 1,
+                "use_image_id": False,
+            }
+        }
+        mode = self.input_mode
+        if mode == "audio":
+            # Fall back to whole-WAV for audio-only ablations.
+            a = self._get_audio_content(video_id)
+            return ([a] if a else None), extra, "first"
+
+        parts = self._get_minicpm_interleaved_parts(video_id, include_audio=(mode == "all"))
+        if not parts:
+            return None, None, "first"
+        return parts, extra, "first"
+
+    def _get_minicpm_interleaved_parts(
+        self,
+        video_id: str,
+        *,
+        include_audio: bool,
+    ) -> list[dict[str, Any]] | None:
+        """Return cached or freshly extracted MiniCPM-style interleaved OpenAI content parts."""
+        cache_key = f"{video_id}|audio={int(include_audio)}|inline={int(self.inline_local_video)}"
+        cached = self._minicpm_interleave_cache.get(cache_key)
+        if cached is not None:
+            return cached
+
+        self._lazy_ensure_hub_media_dir()
+        video_path = self._resolve_local_video_path(video_id)
+        if video_path is None:
+            logger.warning(
+                "MiniCPM interleave pack requires local video under video_dir=%s for video_id=%r",
+                self.video_dir,
+                video_id,
+            )
+            return None
+        audio_path = self._resolve_local_audio_path(video_id) if include_audio else None
+        if include_audio and audio_path is None:
+            logger.warning(
+                "MiniCPM interleave pack (input_mode=all) missing WAV for video_id=%r under %s",
+                video_id,
+                self.video_dir,
+            )
+            return None
+
+        try:
+            frames, audio_segments = self._extract_minicpm_frame_audio_segments(
+                video_path,
+                audio_path=audio_path,
+                include_audio=include_audio,
+            )
+        except Exception:
+            logger.exception(
+                "Failed MiniCPM 1fps extract for video_id=%r path=%s",
+                video_id,
+                video_path,
+            )
+            return None
+
+        if not frames:
+            logger.warning("No frames extracted for video_id=%r", video_id)
+            return None
+        if include_audio and len(audio_segments) != len(frames):
+            logger.warning(
+                "Frame/audio length mismatch for video_id=%r: frames=%d audios=%d",
+                video_id,
+                len(frames),
+                len(audio_segments),
+            )
+            n = min(len(frames), len(audio_segments))
+            frames = frames[:n]
+            audio_segments = audio_segments[:n]
+
+        cache_root: Path | None = None
+        if not self.inline_local_video:
+            assert self.video_dir is not None
+            cache_root = self.video_dir / ".minicpm_daily_omni_interleave" / video_id
+            cache_root.mkdir(parents=True, exist_ok=True)
+
+        parts: list[dict[str, Any]] = []
+        for i, frame in enumerate(frames):
+            jpeg = _pil_to_jpeg_bytes(frame)
+            if self.inline_local_video:
+                b64 = base64.b64encode(jpeg).decode("ascii")
+                parts.append(
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": f"data:image/jpeg;base64,{b64}"},
+                    }
+                )
+            else:
+                assert cache_root is not None
+                frame_path = cache_root / f"frame_{i:04d}.jpg"
+                if not frame_path.is_file() or frame_path.stat().st_size == 0:
+                    frame_path.write_bytes(jpeg)
+                parts.append(
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": frame_path.expanduser().resolve().as_uri()},
+                    }
+                )
+
+            if include_audio:
+                wav = _numpy_to_wav_bytes(audio_segments[i])
+                if self.inline_local_video:
+                    b64 = base64.b64encode(wav).decode("ascii")
+                    parts.append(
+                        {
+                            "type": "audio_url",
+                            "audio_url": {"url": f"data:audio/wav;base64,{b64}"},
+                        }
+                    )
+                else:
+                    assert cache_root is not None
+                    audio_seg_path = cache_root / f"audio_{i:04d}.wav"
+                    if not audio_seg_path.is_file() or audio_seg_path.stat().st_size == 0:
+                        audio_seg_path.write_bytes(wav)
+                    parts.append(
+                        {
+                            "type": "audio_url",
+                            "audio_url": {"url": audio_seg_path.expanduser().resolve().as_uri()},
+                        }
+                    )
+
+        # File URLs are tiny; cache them so repeated QA on the same video skips re-extract.
+        # Inline base64 must not be retained: ~700 Daily-Omni videos would pin GBs of RSS.
+        if not self.inline_local_video:
+            self._minicpm_interleave_cache[cache_key] = parts
+        logger.debug(
+            "MiniCPM interleave packed video_id=%r frames=%d include_audio=%s parts=%d",
+            video_id,
+            len(frames),
+            include_audio,
+            len(parts),
+        )
+        return parts
+
+    @staticmethod
+    def _extract_minicpm_frame_audio_segments(
+        video_path: Path,
+        *,
+        audio_path: Path | None,
+        include_audio: bool,
+        max_num_frames: int = _MINICPM_MAX_NUM_FRAMES,
+    ) -> tuple[list[Any], list[Any]]:
+        """Port of MiniCPM ``get_video_frame_audio_segments`` (stack_frames=1, 1fps)."""
+        import numpy as np
+        from vllm.multimodal.media.audio import load_audio
+
+        num_video_frames, avg_fps = _probe_video_frames_and_fps(video_path)
+        duration = num_video_frames / avg_fps
+        num_seconds = max(1, int(math.ceil(duration)))
+        second_timestamps = list(range(num_seconds))
+
+        if duration > max_num_frames:
+            timestamps = [round(i * 0.1, 1) for i in range(int(duration / 0.1))]
+            frame_idx = [min(int(ts * avg_fps), num_video_frames - 1) for ts in timestamps]
+            pick = _uniform_sample_indices(len(frame_idx), max_num_frames)
+            frame_idx = [frame_idx[i] for i in pick]
+            timestamps = [timestamps[i] for i in pick]
+        else:
+            frame_idx = [min(int(i * avg_fps), num_video_frames - 1) for i in range(num_seconds)]
+            timestamps = second_timestamps
+
+        frames = _decode_video_frames_rgb(video_path, frame_idx)
+
+        audio_segments: list[Any] = []
+        if include_audio:
+            load_path = str(audio_path) if audio_path is not None else str(video_path)
+            audio_np, sr = load_audio(load_path, sr=_MINICPM_AUDIO_SR, mono=True)
+            for i, start_time in enumerate(timestamps):
+                end_time = timestamps[i + 1] if i < len(timestamps) - 1 else duration
+                start_sample = int(start_time * sr)
+                end_sample = int(end_time * sr)
+                segment = audio_np[start_sample:end_sample]
+                if i == len(timestamps) - 1 and len(segment) < 1600:
+                    segment = np.concatenate([segment, np.zeros(1600 - len(segment), dtype=segment.dtype)])
+                if len(segment) == 0:
+                    segment = np.zeros(1600, dtype=np.float32)
+                audio_segments.append(segment)
+
+        return frames, audio_segments
+
+    def _system_text_for_pack_mode(self) -> str:
+        if self.pack_mode == "minicpm-interleave":
+            return MINICPM_OMNI_SYSTEM_TEXT
+        return DAILY_OMNI_SYSTEM_TEXT
+
     @staticmethod
     def _media_desc_for_official_prompt(mode: DailyOmniInputMode) -> str:
         """``media_desc`` in upstream ``build_conversation``."""
@@ -835,14 +1340,16 @@ class DailyOmniDataset(BenchmarkDataset):
         question: str,
         choice: Any,
     ) -> list[dict[str, Any]]:
-        """Map upstream conversation to OpenAI Chat Completions ``messages`` (video_url / audio_url parts)."""
+        """Map upstream conversation to OpenAI Chat Completions ``messages``."""
         user_text = self._official_daily_omni_user_prompt(question, choice)
         mm_list: list[dict[str, Any]] = mm_payload if isinstance(mm_payload, list) else [mm_payload]
         user_content: list[dict[str, Any]] = [*mm_list, {"type": "text", "text": user_text}]
-        return [
-            {"role": "system", "content": [{"type": "text", "text": DAILY_OMNI_SYSTEM_TEXT}]},
-            {"role": "user", "content": user_content},
-        ]
+        system_text = self._system_text_for_pack_mode()
+        messages: list[dict[str, Any]] = []
+        if system_text:
+            messages.append({"role": "system", "content": [{"type": "text", "text": system_text}]})
+        messages.append({"role": "user", "content": user_content})
+        return messages
 
     def sample_by_task_type(
         self,
@@ -900,6 +1407,7 @@ class DailyOmniDataset(BenchmarkDataset):
             f"dataset_split={self.dataset_split!r}, "
             f"video_dir={self.video_dir!r}, "
             f"input_mode={self.input_mode!r}, "
+            f"pack_mode={self.pack_mode!r}, "
             f"inline_local_video={self.inline_local_video!r}, "
             f"max_duration_seconds={self.max_duration_seconds}, "
             f"random_seed={self.random_seed}"