[Core] Make post-load publication warm-only

Close publication leases inside Host Weight Runtime, return only the publication report, and cover the new terminal mappings. Rename the runtime test module to avoid the pytest basename collision reported in #6436. Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
[Core] Add post-load host weight publication
2026-08-21 12:18:12 +08:00 · 2026-08-21 09:05:34 +08:00
587 changed files with 23512 additions and 36623 deletions
@@ -17,7 +17,7 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-model-executor-$$BUILDKITE_PARALLEL_JOB.xml"

    - label: "Simple · Diffusion Test · Shard %N/%t"
      agent_pool: mi300_1
@@ -29,7 +29,7 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+        - "pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-diffusion-$$BUILDKITE_PARALLEL_JOB.xml"

    - label: "Simple · Engine&Entrypoints Test"
      agent_pool: mi300_1
@@ -38,7 +38,7 @@ steps:
      grade: Blocking
      commands:
        - export VLLM_ROCM_USE_AITER=0
-        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
+        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

    - label: "Simple · Other Test"
      agent_pool: mi300_1
@@ -49,7 +49,7 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py"
+        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

 - group: ":card_index_dividers: Diffusion Test"
  steps:
@@ -17,7 +17,7 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-model-executor-$$BUILDKITE_PARALLEL_JOB.xml"

    - label: "Simple · Diffusion Test"
      agent_pool: mi300_1
@@ -28,7 +28,7 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "pytest -sv tests/diffusion -m 'core_model and cpu'  --ignore=tests/diffusion/cache/test_teacache_extractors.py"
+        - "pytest -sv tests/diffusion -m 'core_model and cpu'  --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

    - label: "Simple · Engine&Entrypoints Test"
      agent_pool: mi300_1
@@ -37,7 +37,7 @@ steps:
      grade: Blocking
      commands:
        - export VLLM_ROCM_USE_AITER=0
-        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
+        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

    - label: "Simple · Other Test"
      agent_pool: mi300_1
@@ -48,7 +48,7 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py"
+        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

 - label: "Custom Pipeline Test"
  agent_pool: mi300_1
@@ -16,11 +16,6 @@
 # At least one of --offline/--online is required, so a job with only one mode runs
 # just that half. Total runtime is bounded by the step's timeout_in_minutes.
 #
-# Coverage + artifact upload run only on scheduled weekly E2E sweeps
-# (WEEKLY=1 and BUILDKITE_BRANCH=main — same gate as bootstrap --e2e).
-# All other invocations (PR merge-test, regular main merge, local) run the
-# same combined pytest as the job used before coverage was added.
-#
 # Exits non-zero if any half or the upload failed.
 set -uo pipefail

@@ -49,17 +44,6 @@ if ((${#OFFLINE[@]} == 0 && ${#ONLINE[@]} == 0)); then
    exit 2
 fi

-_should_run_coverage() {
-    [[ "${WEEKLY:-}" == "1" && "${BUILDKITE_BRANCH:-}" == "main" ]]
-}
-
-if ! _should_run_coverage; then
-    PATHS=("${OFFLINE[@]}" "${ONLINE[@]}")
-    echo "--- pytest (no coverage): ${MODEL_ID}"
-    pytest -s -v "${PATHS[@]}" "${PYTEST_ARGS[@]}"
-    exit $?
-fi
-
 # Two steps can test the same model, and artifacts are scoped to the build (a
 # nightly one carries the ready, merge and nightly tiers at once), so unsuffixed
 # reports would share a path -- a glob download then races to rename onto one file.
@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Render and optionally upload Buildkite pipeline YAML with diff-aware logic.

 Bootstrap mode (``bootstrap-upload-steps.yml``):
@@ -59,12 +58,11 @@ BOOTSTRAP_UPLOAD_IF_KEYS = {
 E2E_GROUP_MARKER = "E2E Test"
 CI_MIRROR_HARDWARES_PATH = ROOT / ".buildkite/common/ci_mirror_hardwares.yml"

-# Bootstrap Buildkite ``if`` expressions.
-# ``*_MAIN_IF``: main + env schedule. ``*_LABEL_IF``: PR label (and/or composed with MAIN).
-# ``*_UPLOAD_IF``: full gate for uploading that child pipeline.
-NIGHTLY_MAIN_IF = 'build.branch == "main" && build.env("NIGHTLY") == "1"'
-NIGHTLY_LABEL_IF = (
-    f"({NIGHTLY_MAIN_IF}) || "
+CUDA_NIGHTLY_ONLY = (
+    '(build.pull_request.labels includes "nightly-test") || (build.branch == "main" && build.env("NIGHTLY") == "1")'
+)
+NPU_NIGHTLY_ONLY = (
+    '(build.branch == "main" && build.env("NIGHTLY") == "1") || '
    '(build.branch != "main" && ('
    'build.pull_request.labels includes "nightly-test" || '
    'build.pull_request.labels includes "omni-test" || '
@@ -73,19 +71,6 @@ NIGHTLY_LABEL_IF = (
    'build.pull_request.labels includes "diffusion-x2v-test"'
    "))"
 )
-WEEKLY_E2E_IF = 'build.branch == "main" && build.env("WEEKLY") == "1"'
-WEEKLY_MAIN_IF = 'build.branch == "main" && (build.env("WEEKLY") == "1" || build.env("NON_CRITICAL") == "1")'
-WEEKLY_LABEL_IF = f'({WEEKLY_MAIN_IF}) || (build.branch != "main" && build.pull_request.labels includes "weekly-test")'
-READY_LABEL_IF = 'build.branch != "main" && build.pull_request.labels includes "ready"'
-MERGE_LABEL_IF = 'build.branch != "main" && build.pull_request.labels includes "merge-test"'
-MERGE_MAIN_IF = (
-    'build.branch == "main" && build.env("NIGHTLY") != "1" && '
-    'build.env("WEEKLY") != "1" && build.env("NON_CRITICAL") != "1"'
-)
-READY_UPLOAD_IF = f"({WEEKLY_E2E_IF}) || ({READY_LABEL_IF})"
-MERGE_UPLOAD_IF = f"({WEEKLY_E2E_IF}) || (({MERGE_MAIN_IF}) || ({MERGE_LABEL_IF}))"
-BOOTSTRAP_DISABLED_IF = "false"
-BOOTSTRAP_ENABLED_IF = "true"


 # --- Logging ---
@@ -117,47 +102,74 @@ 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"'
+    nightly_only = NPU_NIGHTLY_ONLY if platform == "npu" else CUDA_NIGHTLY_ONLY
+
    if platform == "npu":
-        ready_upload = READY_LABEL_IF
-        merge_upload = BOOTSTRAP_DISABLED_IF
-        weekly_label_if = BOOTSTRAP_DISABLED_IF
+        ready_pr = (
+            'build.branch != "main" && ('
+            'build.pull_request.labels includes "npu-test" || '
+            'build.pull_request.labels includes "ready"'
+            ")"
+        )
+        nightly_label_if = nightly_only
+        weekly_label_if = disabled
+        merge_base = disabled
    else:
-        ready_upload = READY_UPLOAD_IF
-        merge_upload = MERGE_UPLOAD_IF
-        weekly_label_if = WEEKLY_LABEL_IF
+        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"'
+        nightly_label_if = (
+            '(build.branch == "main" && build.env("NIGHTLY") == "1") || '
+            '(build.branch != "main" && ('
+            'build.pull_request.labels includes "nightly-test" || '
+            'build.pull_request.labels includes "omni-test" || '
+            'build.pull_request.labels includes "tts-test" || '
+            'build.pull_request.labels includes "diffusion-x2iat-test" || '
+            'build.pull_request.labels includes "diffusion-x2v-test"'
+            "))"
+        )
+        weekly_label_if = (
+            '(build.branch == "main" && build.env("WEEKLY") == "1") || '
+            '(build.branch != "main" && build.pull_request.labels includes "weekly-test")'
+        )
+        merge_base = f"({nightly_main}) || (({merge_main}) || ({merge_pr}))"
+
+    ready_base = f"({nightly_main}) || ({ready_pr})"
+    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; WEEKLY=1 / NON_CRITICAL=1 still run L5.
-        # main+WEEKLY=1 also uploads L2/L3 (those steps then pass --e2e).
-        image_expr = f"({NIGHTLY_MAIN_IF}) || ({WEEKLY_MAIN_IF})" if platform == "cuda" else NIGHTLY_MAIN_IF
-        ready_expr = WEEKLY_E2E_IF if platform == "cuda" else BOOTSTRAP_DISABLED_IF
-        merge_expr = WEEKLY_E2E_IF if platform == "cuda" else BOOTSTRAP_DISABLED_IF
-        nightly_expr = NIGHTLY_MAIN_IF
-        weekly_expr = WEEKLY_MAIN_IF if platform == "cuda" else BOOTSTRAP_DISABLED_IF
+        # NIGHTLY=1 still runs L4 + L2/L3 (--e2e); WEEKLY=1 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
+        nightly_expr = nightly_main
+        weekly_expr = weekly_main if platform == "cuda" else disabled
    elif decision.skip_l2_l3:
        l2_enabled = decision.is_run("npu", "l2") if platform == "npu" else decision.is_run("cuda", "l2")
        l3_enabled = platform == "cuda" and decision.is_run("cuda", "l3")

-        ready_expr = ready_upload if l2_enabled else BOOTSTRAP_DISABLED_IF
-        merge_expr = merge_upload if l3_enabled else BOOTSTRAP_DISABLED_IF
-        nightly_expr = NIGHTLY_LABEL_IF
-        weekly_expr = weekly_label_if if platform == "cuda" else BOOTSTRAP_DISABLED_IF
+        ready_expr = ready_base if l2_enabled else disabled
+        merge_expr = merge_base if l3_enabled else disabled
+        nightly_expr = nightly_label_if
+        weekly_expr = weekly_label_if if platform == "cuda" else disabled

-        image_parts = [f"({NIGHTLY_LABEL_IF})"]
+        image_parts = [f"({nightly_label_if})"]
        if platform == "cuda":
            image_parts.append(f"({weekly_label_if})")
        if l2_enabled:
-            image_parts.insert(0, f"({ready_upload})")
+            image_parts.insert(0, f"({ready_base})")
        if l3_enabled:
-            image_parts.insert(1 if l2_enabled else 0, f"({merge_upload})")
+            image_parts.insert(1 if l2_enabled else 0, f"({merge_base})")
        image_expr = " || ".join(image_parts)
    else:
-        image_expr = BOOTSTRAP_ENABLED_IF
-        ready_expr = ready_upload
-        merge_expr = merge_upload if platform == "cuda" else BOOTSTRAP_DISABLED_IF
-        nightly_expr = NIGHTLY_LABEL_IF
-        weekly_expr = weekly_label_if if platform == "cuda" else BOOTSTRAP_DISABLED_IF
+        image_expr = "true"
+        ready_expr = ready_base
+        merge_expr = merge_base if platform == "cuda" else disabled
+        nightly_expr = nightly_label_if
+        weekly_expr = weekly_label_if if platform == "cuda" else disabled

    return {
        "image": _format_bootstrap_if(image_expr),
@@ -8,28 +8,28 @@ steps:
    agents:
      queue: "cpu_queue_premerge"

-  # L2 Test — PR ready label (diff-aware) or main+WEEKLY=1 (--e2e, full E2E group)
+  # L2 Test — PR ready label (diff-aware) or main+NIGHTLY=1 (--e2e, full E2E group)
  - label: "Upload Ready Pipeline"
    depends_on: image-build
    key: upload-ready-pipeline
    commands:
      - |
        EXTRA=""
-        if [ "${WEEKLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; 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+WEEKLY=1 (--e2e, full E2E group)
+  # L3 Test — PR merge-test label (diff-aware) or main+NIGHTLY=1 (--e2e, full E2E group)
  - label: "Upload Merge Pipeline"
    depends_on: image-build
    key: upload-merge-pipeline
    commands:
      - |
        EXTRA=""
-        if [ "${WEEKLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
          EXTRA="--e2e"
        fi
        python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-merge.yml
@@ -45,7 +45,7 @@ steps:
    agents:
      queue: "cpu_queue_premerge"

-  # L5 Test — main+WEEKLY=1 / NON_CRITICAL=1 (scheduled), or PR with weekly-test label
+  # L5 Test — main+WEEKLY=1 (scheduled), or PR with weekly-test label
  - label: "Upload Weekly Pipeline"
    depends_on: image-build
    key: upload-weekly-pipeline
@@ -9,27 +9,23 @@ steps:
    depends_on: upload-merge-pipeline
    steps:
      - label: "Simple · Model Executor Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/model_executor -m 'core_model and cpu'"
+          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

      - label: "Simple · Diffusion Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/diffusion -m 'core_model and cpu'"
+          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

      - label: "Simple · Engine&Entrypoints Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
+          - "timeout 20m pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

      - label: "Simple · Other Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine"
+          - "timeout 20m pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

  - group: ":card_index_dividers: Diffusion Test"
@@ -89,6 +89,32 @@ steps:
            exit $$EXIT
        mirror_hardwares: h100_2

+      - label: ":full_moon: Omni · Perf Test · vLLM Text"
+        key: nightly-omni-performance-vllm-text
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Omni · Perf Test · Multi-Replica"
+        key: nightly-omni-performance-multi-replicas
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_multi_replicas.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_3
+
      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test"
        key: nightly-omni-performance-minicpmo-4-5
        timeout_in_minutes: 180
@@ -229,12 +255,6 @@ steps:
          - pytest -s -v tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py -m full_model --run-level full_model
        mirror_hardwares: h100_4

-      - label: ":full_moon: HiDream-O1-Image · Accuracy Test"
-        timeout_in_minutes: 180
-        commands:
-          - pytest -s -v tests/e2e/accuracy/test_hidream_o1_image.py -m full_model --run-level full_model
-        mirror_hardwares: h100_2
-
      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image"
        key: nightly-diffusion-x2iat-performance-qwen-image
        timeout_in_minutes: 180
@@ -386,6 +406,8 @@ steps:
    depends_on:
      - nightly-omni-performance-no-async-chunk
      - nightly-omni-performance-async-chunk
+      - nightly-omni-performance-vllm-text
+      - nightly-omni-performance-multi-replicas
      - nightly-tts-performance
      - nightly-diffusion-x2iat-performance-qwen-image
      - nightly-diffusion-x2v-performance
@@ -397,6 +419,8 @@ steps:
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-no-async-chunk
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-async-chunk
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-vllm-text
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-multi-replicas
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance
      - python tools/nightly/generate_nightly_perf_excel.py
@@ -9,27 +9,23 @@ steps:
    depends_on: upload-ready-pipeline
    steps:
      - label: "Simple · Model Executor Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/model_executor -m 'core_model and cpu'"
+          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

      - label: "Simple · Diffusion Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/diffusion -m 'core_model and cpu'"
+          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

      - label: "Simple · Engine&Entrypoints Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
+          - "timeout 20m pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

      - label: "Simple · Other Test"
-        timeout_in_minutes: 20
        commands:
-          - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine"
+          - "timeout 20m pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
        mirror_hardwares: l4_1

  - group: ":card_index_dividers: Diffusion Test"
@@ -125,8 +121,32 @@ steps:
          - vllm_omni/outputs/
        commands:
          - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
-        mirror_hardwares: l4_1
-
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              # vLLM 0.27 added a pre-flight capacity check in
+              # distributed/device_communicators/shm_broadcast.py. Without this
+              # the container gets Docker's 64 MiB default and every test in
+              # this step dies at engine start with
+              #   RuntimeError: Insufficient space in /dev/shm:
+              #                 240 MiB required, 64 MiB available
+              # (9 failed in 8.7s on build #2949). vLLM 0.26 had no such check
+              # and allocated regardless, so this step passed on main while
+              # being under-provisioned the whole time — 0.27 turned a silent
+              # risk into an explicit error.
+              #
+              # This is the only inline docker plugin block in .buildkite/cuda/;
+              # every other step inherits common/ci_mirror_hardwares.yml, which
+              # already sets 8gb. Matching that rather than inventing a value.
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"

      - label: "RLHF · VeRL-Omni E2E Test"
        source_file_dependencies:
@@ -5,91 +5,51 @@ env:
  HF_HUB_ETAG_TIMEOUT: 60

 steps:
-  - label: "Simple · CPU Coverage Test"
-    key: weekly-simple-cpu-coverage
-    depends_on: upload-weekly-pipeline
-    if: build.env("WEEKLY") == "1"
-    timeout_in_minutes: 60
-    commands:
-      - |
-        set +e
-        REPORT="coverage-core-model-cpu-$${BUILDKITE_STEP_ID:-local}.xml"
-        pytest -sv tests/ -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:$${REPORT}
-        EXIT=$$?
-        gzip -9 -f "$${REPORT}"
-        buildkite-agent artifact upload "$${REPORT}.gz"
-        exit $$EXIT
-    mirror_hardwares: l4_1

-  - group: "Reliability Test"
+  - label: "Reliability Test - qwen3-omni"
+    timeout_in_minutes: 180
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    commands:
+      - pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
+    mirror_hardwares: h100_2
+
+  - label: "Reliability Test - wan22"
+    timeout_in_minutes: 180
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    commands:
+      - pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
+    mirror_hardwares: h100_2
+
+  - label: "Reliability Test - hunyuan-image"
+    timeout_in_minutes: 180
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    commands:
+      - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
+    mirror_hardwares: h100_4
+
+  - group: "Reliability Test - Invalid parameters Test"
    depends_on: upload-weekly-pipeline
    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
    steps:

-      - label: "Reliability Test · qwen3-omni"
-        timeout_in_minutes: 180
-        commands:
-          - pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
-        mirror_hardwares: h100_2
-
-      - label: "Reliability Test · wan22"
-        timeout_in_minutes: 180
-        commands:
-          - pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
-        mirror_hardwares: h100_2
-
-      - label: "Reliability Test · hunyuan-image"
-        timeout_in_minutes: 180
-        commands:
-          - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
-        mirror_hardwares: h100_4
-
-      - label: "Reliability Test · Invalid parameters · H100"
+      - label: "Invalid parameters Test · H100"
        timeout_in_minutes: 120
        commands:
          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100"
        mirror_hardwares: h100_2

-      - label: "Reliability Test · Invalid parameters · L4"
+      - label: "Invalid parameters Test · L4"
        timeout_in_minutes: 120
        commands:
          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and L4"
        mirror_hardwares: l4_1

-  - group: "Perf Test"
-    depends_on: upload-weekly-pipeline
-    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
-    steps:
-
-      - label: ":full_moon: Omni · Perf Test · vLLM Text"
-        key: weekly-omni-performance-vllm-text
-        timeout_in_minutes: 180
-        commands:
-          - export BENCHMARK_DIR=tests/dfx/perf/results
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
-            exit $$EXIT
-        mirror_hardwares: h100_2
-
-      - label: ":full_moon: Omni · Perf Test · Multi-Replica"
-        key: weekly-omni-performance-multi-replicas
-        timeout_in_minutes: 180
-        commands:
-          - export BENCHMARK_DIR=tests/dfx/perf/results
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_multi_replicas.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
-            exit $$EXIT
-        mirror_hardwares: h100_3
-
  - group: ":card_index_dividers: E2E Tests"
    depends_on: upload-weekly-pipeline
-    if: build.env("NON_CRITICAL") == "1"
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
    steps:

      - label: "Omni · L4"
@@ -115,7 +75,7 @@ steps:
        mirror_hardwares: h100_2

      - label: "TTS · L4"
-        timeout_in_minutes: 120
+        timeout_in_minutes: 90
        commands:
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
          - export VLLM_USE_DEEP_GEMM="0"
@@ -166,7 +126,7 @@ steps:
        mirror_hardwares: l4_4

      - label: "Diffusion · H100 · Single-GPU"
-        timeout_in_minutes: 150
+        timeout_in_minutes: 120
        commands:
          - >-
            pytest -sv
@@ -43,7 +43,7 @@ steps:
    depends_on: upload-ready-pipeline
    steps:
      - label: ":sound: TTS · Qwen3-TTS CustomVoice Test"
-        mirror_hardwares: a3_npu_2
+        mirror_hardwares: a2b3_npu_1
        env:
          VLLM_WORKER_MULTIPROC_METHOD: spawn
          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
@@ -345,8 +345,8 @@ Heavier scenarios in the same file use **`advanced_model` only** (streaming, ext
@pytest.mark.tts
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
-def test_voice_clone_en_non_streaming_001(omni_server, online_client) -> None:
-    online_client.send_audio_speech_request({...})
+def test_voice_clone_en_non_streaming_001(omni_server, openai_client) -> None:
+    openai_client.send_audio_speech_request({...})
 ```

 **L4 consolidation:** Prefer parametrized `OmniServerParams` rows (`default`, `async_chunk`, feature flags) in one expansion module rather than many merge-only files ([#1832](https://github.com/vllm-project/vllm-omni/issues/1832)).
@@ -363,8 +363,8 @@ def test_voice_clone_en_non_streaming_001(omni_server, online_client) -> None:

 | Path | Fixture | Call |
 |------|---------|------|
-| Online `/v1/*` | `online_client` | `online_client.send_*_request(request_config)` |
-| Offline inference | `offline_client` | `offline_client.send_*_request(request_config)` |
+| Online `/v1/*` | `openai_client` | `openai_client.send_*_request(request_config)` |
+| Offline inference | `omni_runner_handler` | `omni_runner_handler.send_*_request(request_config)` |

 1. **Grep `runtime.py` first** — reuse `send_omni_request`, `send_diffusion_request`, `send_audio_speech_request` (online + offline Qwen-style TTS), `send_single_stage_tts_request` (Coqui XTTS / MOSS-TTS-Nano offline), etc.
 2. **No matching helper** → add `send_<feature>_request` (or `send_<route>_http_request` for negative/dfx) in **`runtime.py`** with general `assert_*` bundled inside, **then** call it from the test.
@@ -396,7 +396,7 @@ See [vllm-omni-test skill](../vllm-omni-test/SKILL.md) § **Runtime send helpers
  `/health` return non-200 until you're actually ready.
 - **Mark tests per the CI Levels table** — baseline smoke: `core_model` + `advanced_model`; heavier cases: `advanced_model` only; L4 expansion: `full_model`
 - **No per-model helper modules** — do not create `tests/helpers/{slug}.py`; keep constants and `request_config` payloads in the test file
- **Online and offline e2e go through `runtime.py`** — `online_client.send_audio_speech_request` (online); `offline_client.send_audio_speech_request` (Qwen-style offline) or `send_single_stage_tts_request` (single-stage offline). Add a new `send_*_request` in `runtime.py` when none fits; do not embed `omni.generate` or HTTP in tests
+- **Online and offline e2e go through `runtime.py`** — `openai_client.send_audio_speech_request` (online); `omni_runner_handler.send_audio_speech_request` (Qwen-style offline) or `send_single_stage_tts_request` (single-stage offline). Add a new `send_*_request` in `runtime.py` when none fits; do not embed `omni.generate` or HTTP in tests

 ## Phase 4: Async Chunk (Streaming)

@@ -1,135 +0,0 @@
---
-name: find-simplifications
-description: Find evidence-backed simplification candidates in vLLM-Omni and, when requested, turn them into focused proposals or code changes. Use for audits of dead, duplicated, speculative, over-generalized, unnecessarily defensive, or hand-rolled code. Use review-pr for ordinary correctness review and diffusion-perf-opt for performance-first optimization.
---
-
-# Find vLLM-Omni Simplifications
-
-Find a small number of well-proven ways to reduce code, state, APIs, or maintenance cost without weakening supported behavior. Treat this as an architecture and ownership audit, not a dead-code search or style cleanup.
-
-The default deliverable is a read-only report. Do not add TODOs, design documents, issues, or code unless the user asks for those changes.
-
-## Establish the Contract
-
-Before judging a candidate:
-
-1. Read the repository instructions supplied by the host.
-2. Read [`docs/design/architecture_overview.md`](../../../docs/design/architecture_overview.md).
-3. Select only the relevant active module and feature documents from [`docs/design/index.md`](../../../docs/design/index.md). Archived design pages are historical context, not current contracts.
-4. Identify the comparison base with `git merge-base HEAD origin/main`; do not assume the current branch represents the intended baseline.
-5. If tests or CI coverage may change, use [`vllm-omni-test`](../vllm-omni-test/SKILL.md) for test placement, markers, and commands.
-
-Separate an implementation accident from an intentional boundary. In particular:
-
- vLLM-Omni extends and sometimes overrides upstream vLLM. A similar upstream implementation is evidence for reuse only after comparing API, lifecycle, and multi-stage semantics.
- Registries, import strings, model configuration, deployment YAML, plugins, and optional dependencies create dynamic consumers that a static reference count can miss.
- Hardware backends, model families, execution modes, and connector implementations may be intentionally parallel. Do not propose collapsing one only because it is inactive in the local environment.
- Stage, replica, request, cache, and transport state can look redundant while representing different owners. Prove that two fields or barriers encode the same fact before merging them.
- A shorter hot path is not a simplification if it adds synchronization, tensor copies, host materialization, recompilation, or a measurable quality/performance regression.
-
-## Strong Candidates
-
-A strong candidate removes a meaningful surface and has a clear replacement or deletion boundary. Examples include:
-
- A public method, configuration option, registry entry, route, message field, or event has no production consumer and no compatibility obligation.
- Tests or documentation are the only consumers of behavior that is not part of a supported contract.
- Multiple layers mirror the same request, lifecycle, cache-validity, readiness, or topology state.
- AR, diffusion, platform, or model-specific implementations duplicate a stable common operation and can share it without hiding performance-critical differences.
- An inherited upstream vLLM primitive is reimplemented locally without an Omni-specific semantic difference.
- Compatibility or fallback branches protect environments the project no longer supports.
- Serialization, IPC, output formatting, or connector code materializes the same payload more than once or converts between equivalent representations.
- A helper, abstraction, package split, or extension point has only one real implementation and adds more indirection than policy.
- Hand-written infrastructure can be replaced by an existing project dependency, Python standard-library feature, PyTorch primitive, or upstream vLLM facility with net deletion and equal observability.
- Tests maintain large mocks, snapshots, or expected data solely for an unused API or representation.
-
-Do not elevate typo fixes, formatting, isolated renames, or vague complexity complaints into simplification candidates. Bundle small cleanups only when they support one substantive removal.
-
-## Survey by Ownership Boundary
-
-For broad audits, cover the largest or riskiest production deltas first. Useful domains are:
-
- `vllm_omni/engine/`: orchestration, stage pools, request state, output convergence, cancellation, and replica lifecycle.
- `vllm_omni/worker/` and AR model execution: upstream vLLM overlays, scheduler/cache state, collective RPC, adapters, and weight loading.
- `vllm_omni/diffusion/`: request/step execution, batching, output materialization, IPC, offload, parallelism, and model pipelines.
- `vllm_omni/entrypoints/`: duplicated validation/defaulting, offline/online parity, streaming, and route compatibility.
- Connectors and distributed code: ownership, copies, barriers, handles, cleanup, and failure propagation across process/device/node boundaries.
- Platform backends: genuine vendor differences versus copied control flow that has drifted.
- Configuration, registries, recipes, and examples: duplicate sources of truth and options that never reach a runtime consumer.
- Tests and CI: redundant fixtures, mocks of obsolete contracts, and expensive coverage that can move to a cheaper level without losing the guarded behavior.
-
-When the user explicitly requests a repository-wide or many-candidate audit and parallel agent work is available and authorized, divide these domains among agents and require the same evidence format from each. Otherwise survey them sequentially.
-
-## Prove or Reject Each Candidate
-
-Use `rg` first, then read every relevant caller and implementation. Search exact Python symbols plus serialized names, route paths, configuration keys, registry strings, YAML values, and CLI spellings.
-
-Classify evidence:
-
- **Production:** `vllm_omni/`, runtime configuration, registries, serving entrypoints, connectors, recipes used for supported deployments, and executable loader paths.
- **Validation/support:** `tests/`, benchmarks, examples, docs, CI, and developer tooling.
- **External/dynamic:** upstream vLLM contracts, third-party plugins, model repositories, optional backends, and string-loaded objects. State what was checked and what remains unknown.
-
-For each candidate, answer:
-
-1. What exact surface would disappear or merge?
-2. Who owns the state or behavior today?
-3. Which production consumers exist, including dynamic consumers?
-4. What supported behavior, compatibility, performance, memory, or observability could change?
-5. What is the smallest validation that would catch a bad simplification?
-6. Is the result net deletion after tests, adapters, migration glue, and documentation are counted?
-
-Reject or downgrade the candidate when a live consumer exists and removal is actually a feature decision; when an active design contract explains the separation; when the change only moves complexity; or when hardware/model evidence is unavailable for a performance-sensitive claim.
-
-## Lifecycle and Data-Movement Audits
-
-For asynchronous or distributed code, map the ownership graph before recommending deletion:
-
-```text
-request admission -> orchestrator -> stage/replica -> worker/device
-  -> connector or IPC boundary -> output aggregation -> terminal cleanup
-```
-
-Associate every lock, event, sentinel, queue, readiness flag, timeout, abort path, and cache marker with an owner and transition. Coalesce mechanisms only when they represent the same state and have the same failure boundary. Preserve separate mechanisms when they protect publication versus rollback, local versus remote ownership, first-terminal-outcome arbitration, or cleanup after partial failure.
-
-For tensors and media, also trace device, process, and ownership transitions. Count GPU-to-CPU copies, shared-memory/object-store handles, serialization, decode/encode steps, and unlink/release responsibility. Prefer eliminating a transfer or representation over merely hiding it behind a helper.
-
-## Dependency Replacements
-
-Prefer facilities already available in the repository before adding a dependency. For a proposed replacement:
-
- Name the exact local behavior it covers and any semantics that remain in glue code.
- Check maintenance, license, optional-dependency boundaries, platform support, and transitive footprint.
- Compare net deletion, test burden, import/startup cost, and runtime behavior.
- Benchmark when the replaced code is on a model, scheduler, connector, serialization, or serving hot path.
-
-A wrapper that retains the same state machine or conversion logic is not a simplification.
-
-## Deliver the Result
-
-Report only candidates supported by concrete evidence. For each one include:
-
- **Surface:** files and symbols involved.
- **Evidence:** production consumers found, dynamic paths checked, and relevant design contract.
- **Simplification:** exactly what to remove, merge, inline, or reuse.
- **Benefit:** deleted API/state/code, reduced data movement, or reduced maintenance/test burden.
- **Risk and validation:** behavior that could change and the narrowest useful tests or measurements.
- **Confidence:** high or medium, with the unresolved evidence named for medium-confidence items.
-
-Also list representative ideas that were rejected when that prevents repeated investigation. Prefer three high-confidence candidates over a long speculative inventory.
-
-If the user asks for a durable design proposal, place it under `docs/design/feature/` and register it in `docs/design/index.md`; use current design-document conventions and keep implementation details proportional to the decision. If the user asks for implementation, make the smallest coherent change, update affected tests/docs, and preserve public compatibility unless the user explicitly accepts a breaking change.
-
-## Folding Work from Another Branch
-
-Diff the source branch against its own merge base with `origin/main`, not against the current feature branch. Port only independently justified candidates. Consolidate overlaps into the candidate with stronger evidence, and do not preserve a candidate count by carrying weaker duplicates.
-
-## Validation
-
-Match validation to the outgoing diff:
-
- Documentation or skill changes: run the repository skill validator when available, check links/paths, search for stale source-specific terminology, and run `git diff --check`.
- Python changes: run the narrowest relevant pytest coverage through the repository environment, then lint/format the changed files.
- Distributed, model, hardware, performance, or accuracy changes: run available CPU/static checks and state the exact GPU/NPU/model gap; never present simulated evidence as device validation.
- Before submission, use [`precheck-pr`](../precheck-pr/SKILL.md) for the branch-level self-review.
-
-When summarizing, distinguish checks that passed from checks that were not runnable, and state the snapshot or branch against which the audit was performed.
@@ -67,8 +67,6 @@ Ask: "Quick mode or full mode?" Then walk the checklist for the detected PR type

 **Also run the Examples-Policy check on every PR** using [references/examples-policy.md](references/examples-policy.md). Inspect only Python paths introduced by the diff. Treat a new model-, checkpoint-, vendor-, or family-specific Python example as blocking; do not report pre-existing example debt.

-**Also run the Simplification pass on every PR** using [find-simplifications](../find-simplifications/SKILL.md). Keep it diff-scoped: inspect added or modified public APIs, state, abstractions, fallback and compatibility paths, data movement, and only the adjacent ownership needed to prove a candidate. Do not expand the pass into a repository-wide audit. Zero candidates is a valid pass; report only evidence-backed opportunities. Mark a candidate as blocking only when it proves a correctness or merge-readiness problem, such as newly unreachable code. Otherwise report it as a warning.
-
 **Also confirm local pre-commit gates** from [docs/contributing/README.md](../../../docs/contributing/README.md#linting). GitHub Actions `SKIP` does not prove they passed. Full new-hook list: Omni SPDX (`vLLM-Omni project`); forbidden imports (stdlib `re`/`base64`, pickle, Hugging Face Hub API, Triton/TileLang); no new `torch.cuda` call sites; shellcheck (macOS/Windows must install the binary); mypy-3.10; test files have CI level + hardware marks; TTS `_tts_model_type` branches in `serving_speech.py` must not increase; markdownlint on `docs/`; Buildkite YAML schema. ✗ for growing `allowed_files` / `ALLOWED_FILES` or raising `MAX_MODEL_TYPE_BRANCHES` without review.

 ### Step 5: Print Report
@@ -84,7 +82,6 @@ Pre-check report for <branch>
  PR title format    ✓
  Code quality       ⚠ 1 broad except, 2 Any hints
  Examples policy    ✗ new model-specific Python example
-  Simplification     ⚠ helper duplicates an existing owner
  PR desc integrity  ✓
  Registry/config    ✓
  Dead code          ⚠ 2 warnings
@@ -31,12 +31,9 @@ self-check.
 - [`diffusion-perf-opt`](diffusion-perf-opt/SKILL.md): guides diffusion model
  performance optimization, including profiling traces, parallel strategies,
  stage timing analysis, and benchmark-driven tuning
- [`find-simplifications`](find-simplifications/SKILL.md): finds evidence-backed
-  opportunities to remove or merge dead, duplicated, speculative,
-  over-generalized, or unnecessarily defensive vLLM-Omni code
 - [`precheck-pr`](precheck-pr/SKILL.md): self-checks a branch before creating a
-  PR by validating title format, dead code, simplification opportunities,
-  accuracy and performance claims, and merge readiness
+  PR by validating title format, dead code, accuracy and performance claims,
+  and merge readiness
 - [`quantization`](quantization/SKILL.md): guides quantization method selection,
  model integration, checkpoint loading, and quality/performance validation
  for vLLM-Omni
@@ -101,7 +101,6 @@ If docs and live code disagree, verify the code/tests and report the drift.
 | [tests-docs-checklist.md](references/checks/tests-docs-checklist.md) | Coverage, CI markers, examples, user docs, or PR evidence need review. |
 | [verification.md](references/checks/verification.md) | Hardware, a server, or a runnable affected path is available for active verification. |
 | [examples-policy.md](../precheck-pr/references/examples-policy.md) | The PR adds, copies, or renames Python under `examples/`; apply the canonical policy shared with `precheck-pr`. |
-| [find-simplifications](../find-simplifications/SKILL.md) | Every review; run a diff-scoped subtraction and simplification pass after correctness blockers. |

 ### Delivery and reviewer coordination

@@ -184,13 +183,9 @@ promotion gate. Candidate or draft rules are questions, not blockers, unless
 current code, tests, or policy enforce them. Inspect both sides of any config,
 registry, serialization, connector, cache, or stage boundary.

-Read and apply [find-simplifications](../find-simplifications/SKILL.md) on every
-review. Constrain it to the diff and the adjacent ownership, callers, or
-consumers needed to prove a candidate. Check whether added or expanded helpers,
-classes, state, fallback and compatibility branches, data movement, or public
-behavior can be deleted, merged, moved, or inlined. Zero candidates is a valid
-result. Do not widen the review into repository backlog or report speculative
-style preferences as simplification findings.
+When a diff adds or expands a helper, class, fallback, compatibility branch, or
+public behavior, run a subtraction pass: remove out-of-scope behavior and check
+whether each new abstraction can be deleted, merged, moved, or inlined.

 ### 6. Verify the changed path

@@ -94,8 +94,8 @@ Existing references: `tests/e2e/offline_inference/test_qwen2_5_omni.py` (L2-styl

 | Scenario | Typical location | Fixtures / runner pattern | Baseline markers & level |
 |----------|------------------|---------------------------|---------------------------|
-| **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `offline_client`. **Function (isolation only):** `omni_runner_function` + `offline_client_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
-| **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `online_client`. **Function (isolation only):** `omni_server_function` + `online_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
+| **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `omni_runner_handler`. **Function (isolation only):** `omni_runner_function` + `omni_runner_handler_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
+| **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `openai_client`. **Function (isolation only):** `omni_server_function` + `openai_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
 | **Documentation / runnable examples** | `tests/examples/offline_inference/`, `tests/examples/online_serving/` | **Offline docs (preferred):** extract Python/Bash blocks from the doc README (e.g. `ReadmeSnippet.extract_readme_snippets`), `pytest.mark.parametrize` each snippet, run via `example_runner.run` with a stable `output_subfolder`. **Online docs:** copy client/request scripts into dedicated tests and keep them in sync with the doc page. | Usually **L4**: `advanced_model`, often `example` plus hardware marks matching the nightly docs-example job (see `.buildkite/cuda/test-nightly.yml`). Full conventions: [docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md) (introduced in [PR #1910](https://github.com/vllm-project/vllm-omni/pull/1910): naming, output directory layout, skip rules, avoid trimming `num_inference_steps` without a strong CI reason). |
 | **Performance / benchmark** | `tests/dfx/perf/tests/*.json` + `run_*_benchmark.py` | JSON or script-driven server + load config; assert explicit metrics / baselines | L4 Perf: JSON `mark` with `full_model` + `omni`/`tts`/`diffusion`; wire `test-nightly.yml` Perf steps |
 | **Invalid parameter / negative HTTP validation** | `tests/dfx/reliability/invalid_param_test/` | Live `omni_server` + low-level `send_*_http_request` with `err_code` / `err_message` | `pytest.mark.slow` + `omni` / `tts` / `diffusion` + `@hardware_marks` (`H100` or `L4`); CI in **`test-weekly.yml`** (not ready/merge/nightly) |
@@ -110,9 +110,9 @@ After choosing offline/online/docs/perf, classify the **product under test** and
 |-----------|------------------------------|----------------------------|----------------------------------------|
 | **What it is** | Multimodal LLM pipeline (thinker/talker/stages; text + vision + audio I/O) | Speech synthesis / voice models | Generative diffusion (noise → image, audio, text, or video) |
 | **Examples** | Qwen2.5-Omni, Qwen3-Omni | Qwen3-TTS, VoxCPM2, Higgs-Audio, Step-Audio2 | Qwen-Image, BAGEL, Wan2.2, HunyuanVideo |
-| **Offline runner** | `omni_runner` + `offline_client`, `generate_multimodal` | `Omni(...).generate` with deploy YAML + TTS params | `Omni(...).generate` + `OmniDiffusionSamplingParams` |
-| **Online client** | `online_client.send_omni_request` (chat completions, modalities) | `online_client.send_audio_speech_request` (`/v1/audio/speech`) | `send_diffusion_request` (chat/T2I), `send_video_diffusion_request` (`/v1/videos`, X2V), or `send_images_generations_http_request` / `send_images_edits_http_request` (DALL-E routes) |
-| **Typical assertions** | Stage outputs, text/audio keywords via `OfflineOmniClient` / response handler | WAV bytes, stream chunks, speech endpoint contract | Image/video dimensions, `final_output_type`, `assert_diffusion_response` |
+| **Offline runner** | `omni_runner` + `omni_runner_handler`, `generate_multimodal` | `Omni(...).generate` with deploy YAML + TTS params | `Omni(...).generate` + `OmniDiffusionSamplingParams` |
+| **Online client** | `openai_client.send_omni_request` (chat completions, modalities) | `openai_client.send_audio_speech_request` (`/v1/audio/speech`) | `send_diffusion_request` (chat/T2I), `send_video_diffusion_request` (`/v1/videos`, X2V), or `send_images_generations_http_request` / `send_images_edits_http_request` (DALL-E routes) |
+| **Typical assertions** | Stage outputs, text/audio keywords via `OmniRunnerHandler` / response handler | WAV bytes, stream chunks, speech endpoint contract | Image/video dimensions, `final_output_type`, `assert_diffusion_response` |
 | **Deploy YAML** | Per-model deploy configs (`ci/qwen3_omni_moe.yaml`, …) | `qwen3_tts.yaml`, `voxcpm2.yaml`, … | Often default serve; parallel/offload YAML for heavy DiT |
 | **Nightly group** (`test-nightly.yml`) | **Omni Model Test** — `-m "full_model and omni"` | **TTS Model Test** — `-m "full_model and tts"` | **Diffusion X2I(&A&T)** *or* **Diffusion X2V** (see Step 3 table; same `diffusion` marker, different YAML group / file shard) |
 | **L4 pressure** | Expansion per modality/model as needed | Expansion + accuracy/perf in TTS group | X2I: merge feature combos per [#1832](https://github.com/vllm-project/vllm-omni/issues/1832); X2V: separate nightly group |
@@ -175,9 +175,9 @@ See **L1 unit test constraints (mocking)** below for the full do/don't list.
@pytest.mark.omni
@hardware_test(...)
@pytest.mark.parametrize("omni_runner", test_params, indirect=True)
-def test_<scenario>(omni_runner, offline_client) -> None:
+def test_<scenario>(omni_runner, omni_runner_handler) -> None:
    request_config = {"prompts": ..., "modalities": [...]}  # optional: images, videos, audios
-    offline_client.send_omni_request(request_config)
+    omni_runner_handler.send_omni_request(request_config)
 ```

 *Offline generative e2e — **Diffusion** (representative):*
@@ -187,8 +187,8 @@ def test_<scenario>(omni_runner, offline_client) -> None:
@pytest.mark.diffusion
@hardware_test(...)
@pytest.mark.parametrize("omni_runner", test_params, indirect=True)
-def test_text_to_image_001(offline_client) -> None:
-    offline_client.send_diffusion_request({"prompt": "...", "extra_body": {"num_inference_steps": 4, ...}})
+def test_text_to_image_001(omni_runner_handler) -> None:
+    omni_runner_handler.send_diffusion_request({"prompt": "...", "extra_body": {"num_inference_steps": 4, ...}})
 ```

 *Offline TTS e2e — **Qwen3-TTS** (two-stage; representative):*
@@ -198,8 +198,8 @@ def test_text_to_image_001(offline_client) -> None:
@pytest.mark.tts
@hardware_test(...)
@pytest.mark.parametrize("omni_runner", tts_server_params, indirect=True)
-def test_text_to_audio_001(omni_runner, offline_client) -> None:
-    offline_client.send_audio_speech_request({
+def test_text_to_audio_001(omni_runner, omni_runner_handler) -> None:
+    omni_runner_handler.send_audio_speech_request({
        "input": "...",
        "task_type": "Base",
        "ref_audio": REF_AUDIO_URL,
@@ -214,8 +214,8 @@ def test_text_to_audio_001(omni_runner, offline_client) -> None:
@pytest.mark.tts
@hardware_test(...)
@pytest.mark.parametrize("omni_runner", tts_server_params, indirect=True)
-def test_voice_clone_001(offline_client) -> None:
-    offline_client.send_single_stage_tts_request({
+def test_voice_clone_001(omni_runner_handler) -> None:
+    omni_runner_handler.send_single_stage_tts_request({
        "input": "...",
        "language": "en",
        "prompt_audio_path": REF_AUDIO_PATH,
@@ -231,9 +231,9 @@ def test_voice_clone_001(offline_client) -> None:
@pytest.mark.omni
@hardware_test(...)
@pytest.mark.parametrize("omni_server", test_params, indirect=True)
-def test_text_to_text_001(omni_server, online_client) -> None:
+def test_text_to_text_001(omni_server, openai_client) -> None:
    request_config = {"model": omni_server.model, "messages": ..., "modalities": ["text"]}
-    online_client.send_omni_request(request_config)
+    openai_client.send_omni_request(request_config)
 ```

 *Online serving e2e — **TTS** (representative):*
@@ -244,9 +244,9 @@ def test_text_to_text_001(omni_server, online_client) -> None:
@pytest.mark.tts
@hardware_test(...)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
-def test_text_to_audio_001(omni_server, online_client) -> None:
+def test_text_to_audio_001(omni_server, openai_client) -> None:
    request_config = {"model": omni_server.model, "input": "...", "response_format": "wav", ...}
-    online_client.send_audio_speech_request(request_config)
+    openai_client.send_audio_speech_request(request_config)
 ```

 *Online serving e2e — **Diffusion** X2I (representative):*
@@ -255,8 +255,8 @@ def test_text_to_audio_001(omni_server, online_client) -> None:
@pytest.mark.core_model
@pytest.mark.diffusion
@pytest.mark.parametrize("omni_server", _get_default_case(MODEL), indirect=True)
-def test_text_to_image_001(omni_server, online_client) -> None:
-    online_client.send_diffusion_request({...})  # chat completions + extra_body
+def test_text_to_image_001(omni_server, openai_client) -> None:
+    openai_client.send_diffusion_request({...})  # chat completions + extra_body
 ```

 *Online serving e2e — **Diffusion** X2V (representative):*
@@ -264,8 +264,8 @@ def test_text_to_image_001(omni_server, online_client) -> None:
 ```python
@pytest.mark.core_model
@pytest.mark.diffusion
-def test_text_to_video_001(omni_server, online_client) -> None:
-    online_client.send_video_diffusion_request({"model": ..., "form_data": {...}})  # /v1/videos
+def test_text_to_video_001(omni_server, openai_client) -> None:
+    openai_client.send_video_diffusion_request({"model": ..., "form_data": {...}})  # /v1/videos
 ```

 *Documentation example tests:* follow the **Preferred Test Strategy** in [l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md): dynamic extraction for offline READMEs; explicit copied client code for online pages until extraction is justified; use the documented **naming**, **output directory** (page folder + case id), and **skipping** rules (e.g. Gradio-only scripts).
@@ -324,21 +324,21 @@ vLLM-Omni e2e tests start a real **OmniServer** (online) or **OmniRunner** (offl

 | Scope | Online fixtures | Offline fixtures | When to use |
 |-------|-----------------|------------------|-------------|
-| **Module** (default) | `omni_server` → `online_client` | `omni_runner` → `offline_client` | **Default for L2/L3/L4 expansion** — amortize model/server init across `test_*` in the same module. Same `OmniServerParams` / runner config can be reused by multiple tests. |
-| **Function** | `omni_server_function` → `online_client_function` | `omni_runner_function` → `offline_client_function` | **Only when required** — each `test_*` must get a **clean** server/runner (no shared engine/GPU state). Typical: `tests/dfx/reliability/`, sleep/wakeup, crash/restart, tests that mutate global server state. |
+| **Module** (default) | `omni_server` → `openai_client` | `omni_runner` → `omni_runner_handler` | **Default for L2/L3/L4 expansion** — amortize model/server init across `test_*` in the same module. Same `OmniServerParams` / runner config can be reused by multiple tests. |
+| **Function** | `omni_server_function` → `openai_client_function` | `omni_runner_function` → `omni_runner_handler_function` | **Only when required** — each `test_*` must get a **clean** server/runner (no shared engine/GPU state). Typical: `tests/dfx/reliability/`, sleep/wakeup, crash/restart, tests that mutate global server state. |

 **Rules:**

 1. **Default to module scope** (`omni_server` / `omni_runner`) unless the scenario **explicitly** needs a fresh instance per test function.
-2. **Indirect parametrize name must match the fixture name**: `@pytest.mark.parametrize("omni_server", ...)` with `omni_server` + `online_client`; `@pytest.mark.parametrize("omni_server_function", ...)` with `omni_server_function` + `online_client_function`. Do not mix module fixture with function client (or vice versa).
+2. **Indirect parametrize name must match the fixture name**: `@pytest.mark.parametrize("omni_server", ...)` with `omni_server` + `openai_client`; `@pytest.mark.parametrize("omni_server_function", ...)` with `omni_server_function` + `openai_client_function`. Do not mix module fixture with function client (or vice versa).
 3. Different `OmniServerParams` per `test_*` (e.g. default vs `--enable-cpu-offload`) is still OK with **module** `omni_server` — pytest parametrizes per test node; only switch to `_function` when isolation between tests matters, not merely because `server_args` differ.
 4. One `test_*` with **many** server configs (expansion matrix) → single function + `@pytest.mark.parametrize("omni_server", [...], indirect=True)` + module `omni_server` (see in-tree `test_qwen_image_expansion.py`).

 ```python
 # Default — module-scoped server (L4 expansion)
@pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=H100)], indirect=True)
-def test_foo_images_generations_default_1024(omni_server, online_client) -> None:
-    online_client.send_images_generations_request({...})
+def test_foo_images_generations_default_1024(omni_server, openai_client) -> None:
+    openai_client.send_images_generations_request({...})

 # Function-scoped — reliability / per-test clean state only
@pytest.mark.parametrize(
@@ -346,9 +346,9 @@ def test_foo_images_generations_default_1024(omni_server, online_client) -> None
    [pytest.param(OmniServerParams(model=MODEL), marks=H100)],
    indirect=True,
 )
-def test_foo_sleep_wakeup_cycle(omni_server_function, online_client_function) -> None:
-    online_client_function.send_omni_sleep_http_request({...})
-    online_client_function.send_omni_wakeup_http_request({...})
+def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) -> None:
+    openai_client_function.send_omni_sleep_http_request({...})
+    openai_client_function.send_omni_wakeup_http_request({...})
 ```

 #### Runtime send helpers — online **and** offline (`tests/helpers/runtime.py`)
@@ -357,19 +357,19 @@ def test_foo_sleep_wakeup_cycle(omni_server_function, online_client_function) ->

 | Principle | Action |
 |-----------|--------|
-| **Reuse first** | Grep `runtime.py` for an existing **`send_*_request`** on `OnlineOmniClient` (online) or `OfflineOmniClient` (offline) that matches the endpoint / pipeline shape. |
+| **Reuse first** | Grep `runtime.py` for an existing **`send_*_request`** on `OpenAIClientHandler` (online) or `OmniRunnerHandler` (offline) that matches the endpoint / pipeline shape. |
 | **Extend when close** | If an existing `send_*_request` almost fits (e.g. missing one optional field), extend it in `runtime.py` — do not fork logic in the test file. |
 | **Add when missing** | No suitable helper → add **`send_<feature>_request`** (high-level: call + general `assert_*`) or **`send_<route>_<verb>_http_request`** (low-level HTTP for negative/dfx) in **`runtime.py` first**, then call it from tests. |
 | **Test module owns payload only** | In `test_*.py`: `MODEL`, deploy path, vendored media, `get_prompt()`, and inline **`request_config` dicts** only. **No** `omni.generate(...)`, raw `requests.post`, OpenAI SDK calls, or `_collect_audio()` / `_process_output()` in e2e tests. |

-**Online** (`online_client` from `omni_server`): `OnlineOmniClient.send_*_request`.
+**Online** (`openai_client` from `omni_server`): `OpenAIClientHandler.send_*_request`.

-**Offline** (`offline_client` from `omni_runner`): `OfflineOmniClient.send_*_request`.
+**Offline** (`omni_runner_handler` from `omni_runner`): `OmniRunnerHandler.send_*_request`.

 | Do | Don't |
 |----|-------|
-| Online: `online_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_video_diffusion_request`, `send_images_generations_request`, … | `requests.post(f"{base_url}/v1/...", json=…)` or `client.chat.completions.create(...)` inside a test |
-| Offline: `offline_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_single_stage_tts_request`, `send_single_stage_tts_batch_request`, … | `omni_runner.omni.generate(...)` + hand-rolled tensor/WAV extraction in `test_*.py` |
+| Online: `openai_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_video_diffusion_request`, `send_images_generations_request`, … | `requests.post(f"{base_url}/v1/...", json=…)` or `client.chat.completions.create(...)` inside a test |
+| Offline: `omni_runner_handler.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_single_stage_tts_request`, `send_single_stage_tts_batch_request`, … | `omni_runner.omni.generate(...)` + hand-rolled tensor/WAV extraction in `test_*.py` |
 | Add missing `send_*` to **`runtime.py`** first; bundle general `assert_*` inside the send helper | A one-off `def _post_*` / `def _collect_audio` at the bottom of a test module |
 | Mirror naming/style of neighboring `send_*` (docstring, `request_config` dict, optional `run_level`, `err_code` / `err_message` for negative cases) | Different parameter shapes per test file for the same endpoint |

@@ -392,7 +392,7 @@ def test_foo_sleep_wakeup_cycle(omni_server_function, online_client_function) ->

 Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not** special validation — the general `assert_*` should read those from `request_config` inside `send_*`.

-**`send_*` ↔ `assert_*` pairing (online `OnlineOmniClient`):**
+**`send_*` ↔ `assert_*` pairing (online `OpenAIClientHandler`):**

 | High-level `send_*` (prefer in L2+ e2e) | Assert already invoked inside `runtime.py` | Low-level HTTP-only sibling |
 |----------------------------------------|--------------------------------------------|-----------------------------|
@@ -405,7 +405,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**

 **Rule:** Tests call high-level `send_*_request` for the general contract. **Never** call the same bundled `assert_*` again. Call an extra `assert_*` in the test **only** for special, case-specific validation.

-**Common `OnlineOmniClient` entry points** (non-exhaustive — grep `runtime.py` before adding):
+**Common `OpenAIClientHandler` entry points** (non-exhaustive — grep `runtime.py` before adding):

 | Area | High-level (SDK + assert) | Low-level HTTP (`*_http_request`) |
 |------|---------------------------|-----------------------------------|
@@ -416,7 +416,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 | TTS (online) | `send_audio_speech_request` | `send_audio_speech_http_request`, `send_audio_generate_http_request`, … |
 | Ops / meta | — | `send_health_http_request`, `send_models_http_request`, `send_omni_sleep_http_request`, … |

-**Common `OfflineOmniClient` entry points** (offline — grep `runtime.py` before adding):
+**Common `OmniRunnerHandler` entry points** (offline — grep `runtime.py` before adding):

 | Area | High-level (`send_*_request` + assert) | Notes |
 |------|----------------------------------------|-------|
@@ -426,7 +426,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 | Single-stage TTS (Coqui XTTS, MOSS-TTS-Nano, …) | `send_single_stage_tts_request`, `send_single_stage_tts_batch_request` | `prompt` + `additional_information` + `omni.generate` — **do not** duplicate in tests |
 | New model family | **Add** `send_<family>_request` here first | Then call from `tests/e2e/offline_inference/test_*.py` |

-L1 tests under `tests/entrypoints/` may use **FastAPI `TestClient`** or direct handler calls with mocks; they do **not** need `OnlineOmniClient` / `OfflineOmniClient`. **L2+ online and offline e2e** must use `runtime.py` `send_*` helpers.
+L1 tests under `tests/entrypoints/` may use **FastAPI `TestClient`** or direct handler calls with mocks; they do **not** need `OpenAIClientHandler` / `OmniRunnerHandler`. **L2+ online and offline e2e** must use `runtime.py` `send_*` helpers.

 #### Invalid parameter validation (`tests/dfx/reliability/invalid_param_test/`)

@@ -452,7 +452,7 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
 | **Module markers** | `pytestmark = [pytest.mark.slow, pytest.mark.<omni or tts or diffusion>]` |
 | **Server fixture** | `_PARAMS` / `_QWEN3_TTS_SPEECH`-style list of `pytest.param(OmniServerParams(...), id="...", marks=hardware_marks(...))`; `@pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)` |
 | **Hardware** | `hardware_marks(res={"cuda": "H100"})` for heavy diffusion/omni/video; `hardware_marks(res={"cuda": "L4"})` for smaller TTS models (must match weekly `-m "slow and L4"` step) |
-| **HTTP client** | **Low-level** `online_client.send_*_http_request({..., "err_code": 400, "err_message": (...)} )` — **not** `send_*_request` (success path) |
+| **HTTP client** | **Low-level** `openai_client.send_*_http_request({..., "err_code": 400, "err_message": (...)} )` — **not** `send_*_request` (success path) |
 | **Case shape** | Prefer **one `test_*` per route family** + `@pytest.mark.parametrize("body_spec, err_message", [...])` with stable `id=` per case; or dedicated `test_<route>_malformed_json` when not parametrized |
 | **Body helpers** | `_minimal_<endpoint>_json(omni_server)` / `_minimal_*_form_data()`; merge overrides with `body.update(body_spec)` |
 | **Known gaps** | `pytest.mark.skip(reason="…#3649")` as `_SKIP_ISSUE_3649` when server validation is not yet strict (mirror neighboring cases) |
@@ -485,13 +485,13 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
@pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
 def test_images_generations_invalid_requests(
    omni_server: OmniServer,
-    online_client: OnlineOmniClient,
+    openai_client: OpenAIClientHandler,
    body_spec: dict[str, object],
    err_message: str | tuple[str, ...],
 ) -> None:
    body = _minimal_images_gen_json(omni_server)
    body.update(body_spec)
-    online_client.send_images_generations_http_request(
+    openai_client.send_images_generations_http_request(
        {"json": body, "timeout": 300, "err_code": 400, "err_message": err_message}
    )
 ```
@@ -542,18 +542,18 @@ def send_images_generations_request(self, request_config: dict[str, Any], ...) -

 # tests/e2e/.../test_foo_expansion.py — one case per test_* (no case_id branching)
@pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_images_generations_default_1024(omni_server, online_client) -> None:
-    online_client.send_images_generations_request({"json": body, "timeout": 300})
+def test_foo_images_generations_default_1024(omni_server, openai_client) -> None:
+    openai_client.send_images_generations_request({"json": body, "timeout": 300})

@pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_images_generations_sizes_256_512_1024(omni_server, online_client) -> None:
+def test_foo_images_generations_sizes_256_512_1024(omni_server, openai_client) -> None:
    for size in ("256x256", "512x512", "1024x1024"):
-        online_client.send_images_generations_request({"json": {**body, "size": size}, "timeout": 300})
+        openai_client.send_images_generations_request({"json": {**body, "size": size}, "timeout": 300})

 # chat diffusion — separate test; send_diffusion_request bundles assert_diffusion_response
@pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_chat_completions_t2i_fallback(omni_server, online_client) -> None:
-    online_client.send_diffusion_request({...})
+def test_foo_chat_completions_t2i_fallback(omni_server, openai_client) -> None:
+    openai_client.send_diffusion_request({...})
 ```

 L1 tests under `tests/entrypoints/` may keep **minimal** asserts next to the handler under test when validating a single branch; still prefer `assertions.py` when the same JSON/media check appears in more than one file or level (e.g. L1 protocol test + L4 e2e).
@@ -279,7 +279,7 @@ When adding or modifying tests, do not stop at “where the file lives” — al
 4. **API client + assert placement**: **General** validation → implement in `assertions.py`, call **inside** `send_*_request` in `runtime.py`; tests call **`send_*_request` only**. **Special** case validation → `assert_*` in `assertions.py`, called **in the test** after `send_*_request`. Low-level `send_*_http_request` is for negative/dfx tests (`err_code`), not ordinary L2+ success e2e.
 5. **Shared assertions**: logic in **`tests/helpers/assertions.py`**; general checks inside `send_*_request`; special checks only in the test. No `_assert_*` in `test_*.py`.
 6. **One case → one `test_*`**: function name reflects what is validated; no `if case_id == ...` mega-test merging multiple scenarios.
-7. **Fixture scope**: default **`omni_server` + `online_client`** / **`omni_runner` + `offline_client`** (module). Use **`omni_server_function` + `online_client_function`** / **`omni_runner_function` + `offline_client_function`** only when each `test_*` must spawn a fresh instance. Parametrize name must match fixture (`omni_server` vs `omni_server_function`).
+7. **Fixture scope**: default **`omni_server` + `openai_client`** / **`omni_runner` + `omni_runner_handler`** (module). Use **`omni_server_function` + `openai_client_function`** / **`omni_runner_function` + `omni_runner_handler_function`** only when each `test_*` must spawn a fresh instance. Parametrize name must match fixture (`omni_server` vs `omni_server_function`).
 8. **Type marker**: `omni`, `tts`, or `diffusion` on every model e2e module.
 9. **Diffusion L4 Function**: wire `*_expansion.py` into **X2I(&A&T)** or **X2V** **Function Test** in **`test-nightly.yml` only** — do not add `test-merge.yml` unless the user also requested L3.
 10. **Diffusion L4 Perf** (only when requested): add `tests/dfx/perf/tests/test_<slug>_vllm_omni.json` + **Perf Test · &lt;Model&gt;** step (artifact upload); not part of “L4 functional cases” by default.
@@ -245,6 +245,7 @@ datasets/
 !recipes/cosmos3/negative_prompt.json
 !tests/dfx/perf/tests/*.json
 !apps/ComfyUI-vLLM-Omni/example_workflows/*.json
+!tests/assets/soulxsinger/*.json
 *.jsonl
 *.parquet

@@ -12,10 +12,10 @@ Easy, fast, and cheap omni-modality model serving for everyone
 | <a href="https://vllm-omni.readthedocs.io/en/latest/"><b>Documentation</b></a> | <a href="https://deepwiki.com/vllm-project/vllm-omni"><b>DeepWiki</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> | <a href="docs/assets/WeChat.jpg"><b>WeChat</b></a> | <a href="https://arxiv.org/abs/2602.02204"><b>Paper</b></a> | <a href="https://docs.google.com/presentation/d/1aPj0OGl_-ZVoib-Qne5dGDAlrRFB-PdHl6E-EE99g8E/edit?usp=sharing"><b>Slides</b></a> |
 </p>

+
 ---

 *Latest News* 🔥
-
 - [2026/08] [VeRL-Omni](https://github.com/verl-project/verl-omni) `v0.2.0` is released: faster diffusion RL powered by vLLM-Omni (request-level/step-wise batching with FA3), rebuilt Qwen3-Omni multimodal training (DPO & GSPO), plus LTX-2.3, Qwen-Image-Edit support and more. See the [release notes](https://github.com/verl-project/verl-omni/releases/tag/v0.2.0).
 - [2026/08] We released [0.26.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.26.0) - aligned with the vLLM 0.26 release line, featuring [MiniMax H3](recipes/MiniMaxAI/MiniMax-H3.md) joint video/audio generation, an experimental full-duplex realtime runtime for [MiniCPM-o 4.5](recipes/OpenBMB/MiniCPM-o-4_5.md), distributed layerwise diffusion offload, and broader model, hardware, streaming, TTS, and quantization support.
 - [2026/07] We released [0.24.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.24.0) - aligned with the vLLM 0.24 release line, expanding production-ready coverage across TTS, speech, diffusion, image/video generation, and robot-policy serving, with major Omni stage runtime refactoring, diffusion request-level batching, async output materialization, quantization/cache/memory improvements, and broad CUDA/ROCm/XPU/NPU support.
@@ -89,18 +89,11 @@ If you use vLLM-Omni for your research, please cite our [paper](https://arxiv.or
 ```

 ## Join the Community
-
 Feel free to ask questions, provide feedbacks and discuss with fellow users of vLLM-Omni in `#sig-omni` slack channel at [slack.vllm.ai](https://slack.vllm.ai) or vLLM user forum at [discuss.vllm.ai](https://discuss.vllm.ai).

 ## Star History

-<a href="https://www.star-history.com/?repos=vllm-project%2Fvllm-omni&type=date&legend=top-left">
-  <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=vllm-project/vllm-omni&type=date&theme=dark&legend=top-left&sealed_token=ExgLDZJoQEg27Zfhhut2LqN0GYO6Fw2PWLwPE6JYBUp2BgM3hmsYlwaIVopnUEfbRXidQ4nisumrTdKYydiKhy1SZXipw47qY2_tiUDhCpsPXeXtPuEVKVzBwKs3pw0tiHsJgtSfwXx5yjHXck0Y2SblzFWeJYCkTe1WLGTbUAOIETjXXQJjyCGZvKz5" />
-    <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=vllm-project/vllm-omni&type=date&legend=top-left&sealed_token=ExgLDZJoQEg27Zfhhut2LqN0GYO6Fw2PWLwPE6JYBUp2BgM3hmsYlwaIVopnUEfbRXidQ4nisumrTdKYydiKhy1SZXipw47qY2_tiUDhCpsPXeXtPuEVKVzBwKs3pw0tiHsJgtSfwXx5yjHXck0Y2SblzFWeJYCkTe1WLGTbUAOIETjXXQJjyCGZvKz5" />
-    <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=vllm-project/vllm-omni&type=date&legend=top-left&sealed_token=ExgLDZJoQEg27Zfhhut2LqN0GYO6Fw2PWLwPE6JYBUp2BgM3hmsYlwaIVopnUEfbRXidQ4nisumrTdKYydiKhy1SZXipw47qY2_tiUDhCpsPXeXtPuEVKVzBwKs3pw0tiHsJgtSfwXx5yjHXck0Y2SblzFWeJYCkTe1WLGTbUAOIETjXXQJjyCGZvKz5" />
-  </picture>
-</a>
+[![Star History Chart](https://api.star-history.com/svg?repos=vllm-project/vllm-omni&type=date&legend=top-left)](https://www.star-history.com/#vllm-project/vllm-omni&type=date&legend=top-left)

 ## License

@@ -114,6 +114,10 @@ Configuration classes.
 - [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1DecoderDiTConfig][]
 - [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1EncoderConfig][]
 - [vllm_omni.transformers_utils.configs.cosyvoice3.CosyVoice3Config][]
+- [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Config][]
+- [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLConfig][]
+- [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLTextConfig][]
+- [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLVisionConfig][]
 - [vllm_omni.transformers_utils.configs.omnivoice.OmniVoiceConfig][]
 - [vllm_omni.transformers_utils.configs.voxcpm2.VoxCPM2Config][]

@@ -74,10 +74,8 @@ For fields whose deploy default is `null`, the deploy layer contributes no
 override. The effective value may still come from a platform section, an
 explicit CLI or stage override, or the downstream vLLM engine default.

-Note: for the diffusion path, an omitted `distributed_executor_backend` selects
-`uni` on a single GPU (in-process worker, no MessageQueue / `/dev/shm` output
-segments) and `mp` when `num_gpus > 1`. Set `mp` explicitly to keep a worker
-subprocess on one GPU. `ray` / `external_launcher` are not fully supported yet.
+Note: for diffusion path, `distributed_executor_backend` currently defaults to
+`mp`, and `ray` / `external_launcher` are not fully supported yet.

 ### Stage fields

@@ -334,7 +334,7 @@ These PR shapes all trigger skip-ci:
 Skip-ci does **not** apply when the diff also touches product code (for example `vllm_omni/`), or when test files change assertions, imports, fixtures, or other non-skip logic. If the diff cannot be resolved (non-PR branches outside `main`), CI runs as usual.

 !!! note
-    Skipping L2/L3 does **not** disable the Docker image build step. Nightly (L4) upload can still run when the PR has a `nightly-test` label or on scheduled `main` builds with `NIGHTLY=1`. Weekly (L5) upload can still run with `weekly-test`, `WEEKLY=1`, or `NON_CRITICAL=1`. Bootstrap child steps live in `bootstrap-upload-steps.yml`; the hook entry step runs `upload_pipeline.py --upload <platform>/bootstrap-upload-steps.yml`, which injects `if` by step `key` from skip-ci before upload. See [CI Settings — Diff-aware CI](./ci/ci_settings.md#diff-aware-ci) and [Test System Overview](./ci/test_system_overview.md).
+    Skipping L2/L3 does **not** disable the Docker image build step. Nightly (L4) upload can still run when the PR has a `nightly-test` label or on scheduled `main` builds with `NIGHTLY=1`. Bootstrap child steps live in `bootstrap-upload-steps.yml`; the hook entry step runs `upload_pipeline.py --upload <platform>/bootstrap-upload-steps.yml`, which injects `if` by step `key` from skip-ci before upload. See [CI Settings — Diff-aware CI](./ci/ci_settings.md#diff-aware-ci) and [Test System Overview](./ci/test_system_overview.md).

 ### Code Quality

@@ -8,7 +8,7 @@ For CI levels (L1–L5) and triggers, see [Test System Overview](./test_system_o

 Canonical layout (prefer these paths for new changes):

-```text
+```
 .buildkite/
 ├── common/                          # Shared across platforms
 │   ├── scripts/
@@ -48,7 +48,7 @@ Canonical layout (prefer these paths for new changes):
    └── scripts/
 ```

-### Placement rules
+**Placement rules**

 | Rule | Detail |
 | ---- | ------ |
@@ -81,14 +81,7 @@ There are still **legacy copies** at `.buildkite/*.yaml` (without the `cuda/` pr
    | L2 | `cuda/test-ready.yml` | `ready` label |
    | L3 | `cuda/test-merge.yml` | `merge-test` label / main merge |
    | L4 | `cuda/test-nightly.yml` | `nightly-test` label or `NIGHTLY=1` |
-    | L5 | `cuda/test-weekly.yml` | `weekly-test` label, `WEEKLY=1`, or `NON_CRITICAL=1` (see step `if`s below) |
-
-    On scheduled `main` builds, `test-weekly.yml` is uploaded when **`WEEKLY=1` or `NON_CRITICAL=1`**. Inside that file:
-
-    - **Reliability** / **Perf Test** / **Simple · CPU Coverage Test** → `WEEKLY=1` (or PR label `weekly-test` for Reliability/Perf). Pipeline upload for scheduled env vars still requires `main`.
-    - **E2E Tests** group (slow Omni/TTS/Diffusion sweeps) → `NON_CRITICAL=1`
-
-    **`--e2e`:** when uploading ready/merge on `main` with **`WEEKLY=1`**, bootstrap passes `--e2e` so only the **E2E Test** group is kept (same flag `run_cov_split.sh` uses to enable per-model coverage).
+    | L5 | `cuda/test-weekly.yml` | `weekly-test` label or `WEEKLY=1` |

    **Upload:** `upload_pipeline.py --upload` expands uploader-only keys before Buildkite upload.

@@ -119,8 +112,6 @@ There are still **legacy copies** at `.buildkite/*.yaml` (without the `cuda/` pr

    **Test YAML:** `npu/test-npu-ready.yml` (L2), `test-npu-nightly.yml` (L4).

-    **L2 trigger:** PR label `ready` (CUDA/AMD L2). PR labels `ready` + `npu-test` trigger NPU CI only.
-
    **Upload:** same as CUDA—`upload_pipeline.py --upload`.

    **Hardware in YAML:** `mirror_hardwares` preset (string), expanded to `agents`, top-level `image`, and `plugins`. Presets: `a2b3_npu_1`, `a2b3_npu_4`, `a3_npu_2` in `common/ci_mirror_hardwares.yml`.
@@ -213,15 +204,15 @@ Label triggers (`ready`, `merge-test`) are unchanged—diff-aware logic only red
 | Diff shape | Path | Default L2/L3 |
 | --- | --- | --- |
 | Product code / non-whitelisted paths | normal CI | all on |
-| Docs and/or qualifying skip-mark only | `skip_all` | all off (scheduled `main` NIGHTLY/WEEKLY/`NON_CRITICAL` exceptions below) |
+| Docs and/or qualifying skip-mark only | `skip_all` | all off (scheduled `main` NIGHTLY/WEEKLY exceptions below) |
 | Whitelisted CI YAML only | yaml-gated (`skip_l2_l3`) | per platform/level matrix |
 | Docs/skip-mark **+** whitelisted CI YAML | yaml-gated (`skip_l2_l3`) | same as CI-YAML-only (does **not** widen to normal CI) |
 | Non-qualifying skip-mark and no CI YAML | normal CI | all on |
 | Diff unavailable | normal CI | all on |

-**`skip_all` exceptions (CUDA/NPU bootstrap only):** PR labels (`nightly-test`, `merge-test`, `weekly-test`, `ready`, …) do **not** revive jobs under `skip_all`. On `main`, scheduled `NIGHTLY=1` still builds the image and uploads L4; `WEEKLY=1` or `NON_CRITICAL=1` still builds the image and uploads L5 (CUDA). `WEEKLY=1` on `main` also uploads L2/L3 with `--e2e`.
+**`skip_all` exceptions (CUDA/NPU bootstrap only):** PR labels (`nightly-test`, `merge-test`, `npu-test`, `weekly-test`, …) do **not** revive jobs. On `main`, scheduled `NIGHTLY=1` still builds the image and uploads L4 plus L2/L3 (with `--e2e`); `WEEKLY=1` still builds the image and uploads L5 (CUDA).

-**Yaml-gated nightly/weekly:** L4/L5 upload steps keep their normal label / `NIGHTLY` / `WEEKLY` / `NON_CRITICAL` conditions (for example PR `nightly-test` still uploads nightly). Only L2/L3 upload steps are matrix-gated.
+**Yaml-gated nightly/weekly:** L4/L5 upload steps keep their normal label / `NIGHTLY` / `WEEKLY` conditions (for example PR `nightly-test` still uploads nightly). Only L2/L3 upload steps are matrix-gated.

 ##### Whitelisted CI YAML → platform

@@ -294,12 +285,12 @@ CUDA **L2** (`.buildkite/cuda/test-ready.yml`) and **L3** (`.buildkite/cuda/test
 | `source_file_dependencies` | List of path **prefixes**. If any changed file equals a prefix or starts with `prefix/`, keep the step (or group); otherwise omit it. |
 | `mirror_hardwares` | Expand to `agents` + `plugins` (+ optional `image`) from `ci_mirror_hardwares.yml`. |

-### Policy
+**Policy**

 - **Always uploaded** (no key): groups outside **E2E Test**—Simple Test, Diffusion unit tests, Engine/Model Executor, Distributed, Custom Pipeline, Entrypoints (L2), LoRA / Entrypoints (L3).
 - **Diff-gated**: every **E2E Test** leaf job. List the smallest prefix set per step—pytest file(s), model code under `vllm_omni/model_executor/models/` or `vllm_omni/diffusion/models/`, plus `stage_input_processors/` and `vllm_omni/deploy/*.yaml` when applicable. A **group** may define the key instead; the whole group drops if no prefix matches.

-### YAML examples
+**YAML examples**

 ```yaml
      - label: "Diffusion · Qwen Image Test"
@@ -321,7 +312,7 @@ CUDA **L2** (`.buildkite/cuda/test-ready.yml`) and **L3** (`.buildkite/cuda/test
        mirror_hardwares: l4_4
 ```

-### Local dry-run
+**Local dry-run**

 ```bash
 python3 .buildkite/common/scripts/upload_pipeline.py .buildkite/cuda/test-ready.yml
@@ -333,19 +324,18 @@ On PR builds, `upload_pipeline.py` logs `skip '…' (no changes under …)` for

 ### Labels and grouping

- Use **groups** for dashboard readability; keep **E2E Test** as the group name CUDA diff filtering expects for `--e2e` weekly runs on main (`WEEKLY=1`).
+- Use **groups** for dashboard readability; keep **E2E Test** as the group name CUDA diff filtering expects for `--e2e` nightly runs on main.
 - Prefix labels by model domain: **Omni ·**, **TTS ·**, **Diffusion ·**, **Simple ·**, etc., matching existing steps.

 ### Per-model coverage

 Pilot (v1): one Merge-tier (L3) job in `.buildkite/cuda/test-merge.yml` uploads
 per-model, per-entry-mode coverage as Buildkite artifacts — "TTS · Qwen3-TTS Base
-Test". Coverage + artifact upload run only when bootstrap uploaded that job with
-`--e2e` (`WEEKLY=1` and `BUILDKITE_BRANCH=main`, via `run_cov_split.sh`); otherwise
-the same job runs plain pytest without `--cov`. It runs on a single GPU so the
-pilot is cheap to reproduce.
-
-**L1 package coverage (CPU):** ready/merge Simple Test jobs run without `--cov`. Scheduled weekly on `main` with `WEEKLY=1` runs **Simple · CPU Coverage Test** in `.buildkite/cuda/test-weekly.yml` (`pytest -m 'core_model and cpu'` + Cobertura XML artifact).
+Test". It runs on a single GPU so the pilot is cheap to reproduce. (A second
+diffusion pilot, "Diffusion · Bagel Test", was retired from merge; reintroduce
+another diffusion coverage job if a second pilot is needed again.) This is a
+pilot, not a rollout: no dashboard/visualization lives in this repo, and nothing
+is gated on coverage.

 **Naming convention**: `coverage-<model_id>-<mode>-<step_id>.xml.gz`, where
 `<model_id>` is the model's directory name under
@@ -364,8 +354,8 @@ imported. A run covering zero lines produces the same 7 MB. gzip takes it to
 roughly 410 KB and changes nothing about the data.

 `<step_id>` is what keeps two steps that cover the same model apart. Artifacts are
-scoped to the build, and a weekly build can carry ready/merge (with `--e2e`) and
-weekly tiers at once, so several same-model steps land in one namespace — `qwen3_tts` already has
+scoped to the build, and a nightly build carries the ready, merge and nightly tiers
+at once, so several same-model steps land in one namespace — `qwen3_tts` already has
 both a Base and a CustomVoice job, and `minicpmo_4_5` a base and a duplex one. Two
 uploads on one path do not fail loudly: an exact-name download can report the path
 as ambiguous, and a glob download fetches every match concurrently and renames each
@@ -435,7 +425,7 @@ just confirming both files exist.
 | Check | Command / location |
 | ----- | ------------------ |
 | CUDA render | `python3 .buildkite/common/scripts/upload_pipeline.py .buildkite/cuda/test-<level>.yml` |
-| No leaked uploader keys | Pipe the render into `grep -E 'mirror_hardwares\|source_file_dependencies'` — expect empty |
+| No leaked uploader keys | `… \| grep -E 'mirror_hardwares|source_file_dependencies'` → empty |
 | Skip-ci / upload unit tests | `pytest tests/buildkite/` |
 | Local job replay (CUDA L2+) | `tools/run_ready_jobs.sh`, `tools/run_merge_jobs.sh`, `tools/nightly/run_nightly_jobs.sh` (read YAML from `cuda/`) |

@@ -1,8 +1,8 @@
-#### Scope
+**Scope**

 For diffusion models, the L4 functionality test covers all or common *diffusion features* that are supported by this model, including several [parallelism acceleration methods](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/parallelism_acceleration/), [CPU offloading](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/cpu_offload/), [TeaCache](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/teacache/) and [Cache-DiT](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/cache_dit_acceleration/) cache backends, [quantization methods](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/quantization/overview/).

-#### Test Case Design
+**Test Case Design**

 For a *high priority* model (currently listed in [issue #1832](https://github.com/vllm-project/vllm-omni/issues/1832)), we use several test cases, each with multiple features turned on, so that each supported feature is tested in at least one test case. This is to relieve the GPU workload on the CI machine. The suggested test case combination is as follows:

@@ -34,12 +34,12 @@ For a *normal priority* model, further reduce the number of test cases.

 Currently all the features are available in online serving mode. Hence, only need to add `tests/e2e/online_serving/test_{model}_expansion.py`.

-#### Code Style
+**Code Style**

- Validation: test that the multimodal output files of your model have the correct shapes. `OnlineOmniClient.send_diffusion_request` should have taken care of this.
+- Validation: test that the multimodal output files of your model have the correct shapes. `OpenAIClientHandler.send_diffusion_request` should have taken care of this.
 - Test marks: always add `full_model` and `diffusion` for L4 nightly `test_*_expansion.py` cases. Add GPU-related marks if needed. Ref: [Markers for Tests](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/ci/test_writing_guide/#markers-for-tests).
 - To maximize code reuse, you may refer to
-    - `tests/conftest.py` for `omni_server` (running server in subprocess) and `online_client` fixtures (sending requests and validating output), `generate_synthetic_image` and `assert_XXX_valid` helper.
+    - `tests/conftest.py` for `omni_server` (running server in subprocess) and `openai_client` fixtures (sending requests and validating output), `generate_synthetic_image` and `assert_XXX_valid` helper.
    - `tests/helpers/mark.py` for `@hardware_test(...)` and `hardware_marks`.
    - [Parametrizing tests (pytest doc)](https://docs.pytest.org/en/stable/example/parametrize.html) to reuse test function implementation for different cases.
 - Doc: add a concise docstring for each test function.
@@ -1,15 +1,14 @@
 When you want to add L4-level ***performance test*** cases, add entries to JSON files under `tests/dfx/perf/tests/` and run them via `tests/dfx/perf/scripts/run_benchmark.py` (omni / TTS) or `run_diffusion_benchmark.py` (diffusion).

-#### Config file layout (in-tree examples)
+**Config file layout (in-tree examples)**

 | Model type | Runner | Example JSON files |
 | ---------- | ------ | ------------------ |
-| Omni (nightly) | `run_benchmark.py` | `test_qwen3_omni_no_async_chunk.json`, `test_qwen3_omni_async_chunk.json` (`full_model` in `mark`) |
-| Omni (weekly) | `run_benchmark.py` | `test_qwen3_omni_vllm_text.json`, `test_qwen3_omni_multi_replicas.json` (`slow` in `mark`; **Perf Test** in `test-weekly.yml`) |
+| Omni | `run_benchmark.py` | `test_qwen3_omni_no_async_chunk.json`, `test_qwen3_omni_async_chunk.json`, `test_qwen3_omni_vllm_text.json`, `test_qwen3_omni_multi_replicas.json` |
 | TTS | `run_benchmark.py` | `test_tts.json`, `test_voxcpm2.json`, `test_higgs_audio_v3.json` |
 | Diffusion | `run_diffusion_benchmark.py` | `test_qwen_image_vllm_omni.json`, `test_bagel_vllm_omni.json`, `test_wan22_i2v_vllm_omni.json`, `test_cosmos3_vllm_omni.json`, … |

-#### How runners pick cases
+**How runners pick cases**

 Without **`--test-config-file`**, each runner scans all `*.json` under `tests/dfx/perf/tests/` but only keeps its own model type:

@@ -18,9 +17,9 @@ Without **`--test-config-file`**, each runner scans all `*.json` under `tests/df

 Diffusion cases are detected when the JSON has `server_type` (typically `"vllm-omni"`) or `"diffusion"` in the `mark` array. Omni / TTS JSON has neither.

-#### Running perf cases
+**Running perf cases**

-Pass **`--test-config-file`** to run one JSON file as-is, or omit it for the bulk scan above and filter with pytest **`-m`** on each case's JSON `mark` (for example `pytest … run_benchmark.py -m "full_model and tts and H100"` or `-m "slow and omni and H100"` for weekly Omni configs).
+Pass **`--test-config-file`** to run one JSON file as-is, or omit it for the bulk scan above and filter with pytest **`-m`** on each case's JSON `mark` (for example `pytest … run_benchmark.py -m "full_model and tts and H100"`).

 ```JSON
 {
@@ -56,22 +55,20 @@ Pass **`--test-config-file`** to run one JSON file as-is, or omit it for the bul
 }
 ```

-#### Parameter Explanation
+**Parameter Explanation**

-##### Overview
+*Overview*

-| Field              | Required       | Description                                  |
-| ------------------ | -------------- | -------------------------------------------- |
-| test_name          | Yes            | Unique identifier for the test case          |
-| mark               | No             | Pytest marks; see **`mark` field** below     |
-| server_params      | Yes            | Server-side configuration parameters         |
-| benchmark_params   | Yes            | Benchmark running parameters                 |
-| server_type        | Diffusion only | Routes case to run_diffusion_benchmark.py    |
-| benchmark_endpoint | Diffusion only | Benchmark API path                           |
+| Field            | Required | Description                                                     |
+| ---------------- | -------- | --------------------------------------------------------------- |
+| test_name        | Yes      | Unique identifier for the test case                             |
+| mark             | No       | Pytest marks for this case (see **`mark` field** below). Omit only for configs not meant to be filtered by `-m`. |
+| server_params    | Yes      | Server-side configuration parameters                            |
+| benchmark_params | Yes      | Benchmark running parameters (supports multiple configurations) |
+| server_type      | Diffusion only | When set (e.g. `"vllm-omni"`), routes the case to `run_diffusion_benchmark.py`. |
+| benchmark_endpoint | Diffusion only | API path for the benchmark client (e.g. `/v1/videos`, `/v1/images/generations`). |

-Omit `mark` only for configs not meant to be filtered by `-m`. `server_type` is typically `"vllm-omni"`. `benchmark_endpoint` examples: `/v1/videos`, `/v1/images/generations`.
-
-#### `mark` field
+**`mark` field**

 Optional top-level field on each perf JSON **case object** (one per `test_name`). `run_benchmark.py` and `run_diffusion_benchmark.py` read it via `tests.dfx.conftest.resolve_pytest_marks` and attach the marks to the corresponding `pytest.param`, so you can filter locally with `-m`.

@@ -122,11 +119,11 @@ Multi-GPU diffusion (example: Cosmos3 with `cfg-parallel-size=2`):
 - Diffusion perf: include `"diffusion"` in the `mark` array
 - HunyuanImage local-weight cases: add `"local_model"` to the `mark` array

-#### Parametrization IDs
+**Parametrization IDs**

 Each `(server, benchmark index)` pair becomes one `pytest.param` with id `{test_name}-{suffix}`. The suffix comes from `benchmark_params[].name` when set (for example `test_tts-p0`, `test_omni-p1`); otherwise it is derived from fields like `task` / `eval_phase`.

-#### Benchmark result filenames
+**Benchmark result filenames**

 Result files use the **runtime** hardware label from `get_runtime_resource_label()` (detected GPU/NPU name on the machine that ran the job), **not** `mark.hardware_marks.res`. On the default H100 CI pool, `H100` is **omitted** from filenames (`resource_label_for_filename`).

@@ -135,7 +132,7 @@ Examples:
 - Omni/TTS: `result_{test_name}_{optional_hw}_{dataset}_....json` under `BENCHMARK_DIR`
 - Diffusion: one aggregate `diffusion_result_{config_stem}_{optional_hw}_{timestamp}.json` per source JSON file (array of all runs from that file)

-#### Local commands
+**Local commands**

 ```bash
 # Bulk load + filter by JSON mark
@@ -149,19 +146,18 @@ pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
  --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
 ```

-<!-- markdownlint-disable-next-line MD051 -->
 See also [Markers for Tests](#markers-for-tests) for registered hardware markers (`H100`, `L4`, `cuda`, `distributed_cuda`, …).

-#### `server_params` Configuration
+**`server_params` Configuration**

-##### Basic Parameters
+*Basic Parameters*

-| Parameter         | Required | Example                   | Description                    |
-| ----------------- | -------- | ------------------------- | ------------------------------ |
-| model             | Yes      | Qwen/Qwen3-Omni-30B-A3B…  | Model name or path             |
-| stage_config_name | Yes      | qwen3_omni_moe.yaml       | Deploy configuration file name |
+| Parameter         | Required | Example                            | Description                   |
+| ----------------- | -------- | ---------------------------------- | ----------------------------- |
+| model             | Yes      | "Qwen/Qwen3-Omni-30B-A3B-Instruct" | Model name or path            |
+| stage_config_name | Yes      | "qwen3_omni_moe.yaml"              | Deploy configuration file name |

-##### Dynamic Configuration (update/delete)
+*Dynamic Configuration (update/delete)*

 Supports incremental modifications based on the basic configuration:

@@ -170,9 +166,9 @@ Supports incremental modifications based on the basic configuration:
 | update    | Update or add configuration items    |
 | delete    | Delete specified configuration items |

-##### Example
+**Example**:

-```json
+```
 "update": {
    "async_chunk": true,  // Enable asynchronous chunk processing
    "stages": {
@@ -188,22 +184,20 @@ Supports incremental modifications based on the basic configuration:
 }
 ```

-#### `benchmark_params` Configuration
+**`benchmark_params` Configuration**

 You can add any benchmark running parameters you need here. For all optional parameters, refer to the [benchmark documentation](https://github.com/vllm-project/vllm-omni/blob/main/docs/cli/bench/serve.md). General modifications are as follows:

-1. Change the --xxx-xx-xx running parameters to xxx_xx_xx format and fill them as keys in the JSON file.
-2. For boolean variables in the running parameters, modify them to forms such as ignore_eos: true/false and fill them into the JSON file.
-3. Optionally add a `baseline` object (see **Baseline thresholds** below). If you omit `baseline` or leave it empty, the performance test still runs but does not assert metric thresholds from this field.
-4. Set `"name"` on each `benchmark_params` entry for stable pytest ids and readable result keys.
-5. The qps and concurrency modes are recommended to be mutually exclusive. For detailed explanations, see the table below:
+1.  Change the --xxx-xx-xx running parameters to xxx_xx_xx format and fill them as keys in the JSON file.
+2.  For boolean variables in the running parameters, modify them to forms such as ignore_eos: true/false and fill them into the JSON file.
+3.  Optionally add a `baseline` object (see **Baseline thresholds** below). If you omit `baseline` or leave it empty, the performance test still runs but does not assert metric thresholds from this field.
+4.  Set `"name"` on each `benchmark_params` entry for stable pytest ids and readable result keys.
+5.  The qps and concurrency modes are recommended to be mutually exclusive. For detailed explanations, see the table below:

-| Parameter       | Type          | Required | Example/Values       | Description                          |
-| --------------- | ------------- | -------- | -------------------- | ------------------------------------ |
-| name            | string        | No       | `"1024x1024_steps4"` | Stable suffix for pytest param id    |
-| num_prompts     | int / array   | Yes      | 10, [10, 20, 30]     | Number of requests (scalar or sweep) |
-| request_rate    | float / array | No       | 0.5, [0.5, 1, inf]   | Queries per second (QPS mode)        |
-| max_concurrency | int / array   | No       | 1, [1, 2, 3]         | Max in-flight requests               |
-| baseline        | object        | No       | see below            | Optional hardware-nested thresholds  |
-
-`num_prompts`, `request_rate`, and `max_concurrency` accept scalars or arrays; single values expand to match sweep length. QPS and concurrency modes are mutually exclusive. `baseline` top-level keys must be hardware pytest markers from `pyproject.toml`; metric names and values are free-form (scalars or sweep-aligned lists). Omit `baseline` to run without threshold assertions.
+| Parameter       | Type        | Required | Example/Values  | Description                                                                                                                                                                                                                                                          |
+| --------------- | ----------- | -------- | --------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| name            | string      | No       | `"1024x1024_steps4"` | Stable suffix for pytest param id and logs. |
+| 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 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. |
@@ -1,20 +1,13 @@
 # Test Execution Guide
-
 ## Setting Up the Test Environment
-
 ### Creating a Container
-
 vLLM-Omni provides an official Docker image for deployment. These images are built upon vLLM Docker images and are available on [Docker Hub](https://hub.docker.com/r/vllm/vllm-omni/tags). The version of vLLM-Omni indicates which vLLM release it is based on.
 For a local test environment, you can follow the steps below to create a container:
-
 ## Installing Dependencies
-
 ### vLLM & vLLM-Omni
-
 vLLM-Omni is built based on vLLM. You can follow [install guide](../../getting_started/installation/README.md) to build your local environment.

 ### Test Case Dependencies
-
 When running test cases, you may need to install the following dependencies:

 ```bash
@@ -23,7 +16,6 @@ apt-get install -y espeak-ng jq
 ```

 ## Running Tests
-
 Our test scripts use the pytest framework. First, please use `git clone https://github.com/vllm-project/vllm-omni.git` to download the vllm-omni source code. Then, in the root directory of vllm-omni, you can run the following commands in your local test environment to execute the corresponding test cases.

 ### CI job runners (L2–L4): logs, timing, and timeouts
@@ -33,7 +25,7 @@ Our test scripts use the pytest framework. First, please use `git clone https://
 **Log layout** (default under `logs/ready_jobs`, `logs/merge_jobs`, or `logs/nightly_jobs`; override with `--log-dir`):

 | Path | Description |
-| ------ | ------------- |
+|------|-------------|
 | `jobs/*.sh` | Generated per-step bash wrappers |
 | `jobs/.job_timeouts` | Job key → `timeout_in_minutes` from Buildkite YAML (when set) |
 | `<job_key>.log` | Tee output for each job |
@@ -41,7 +33,7 @@ Our test scripts use the pytest framework. First, please use `git clone https://

 **Timing:** while each job runs, stderr prints `finished in …` or `failed after …`. When all jobs finish, a summary is printed and written to `timing_summary.log`, for example:

-```text
+```
 === Job timing summary ===
  TTS_VoxCPM2_Test  25m 40s  OK
  Omni_Qwen3-Omni_Test  8m 05s  TIMED OUT
@@ -122,6 +114,7 @@ Failed jobs: 1/2
    pytest -s -v -m "advanced_model and distributed_cuda and L4" --run-level=advanced_model
    ```

+
 === "L4 level"

    **Recommended:** run CI-aligned nightly jobs from the repo root via [`tools/nightly/run_nightly_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/nightly/run_nightly_jobs.sh). The script reads [`.buildkite/cuda/test-nightly.yml`](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-nightly.yml), generates per-step bash wrappers, runs pytest (perf jobs first, then aggregates perf JSON into Excel), and tees logs under `logs/nightly_jobs/` (see [CI job runners](#ci-job-runners-l2l4-logs-timing-and-timeouts); `generate_nightly_perf_excel` is included in the timing summary). Requires `bash`, `python3`, and PyYAML (`pip install pyyaml`).
@@ -167,7 +160,7 @@ Failed jobs: 1/2
    pytest -sv tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json
    pytest -sv tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
    ```
-    Nightly **Perf Test** jobs in [``test-nightly.yml``](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-nightly.yml) use ``--test-config-file`` only (no ``-m``). Weekly **Perf Test** in [``test-weekly.yml``](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-weekly.yml) runs ``test_qwen3_omni_vllm_text.json`` and ``test_qwen3_omni_multi_replicas.json`` (JSON ``mark`` includes ``slow``). E2e L4 function tests use ``full_model`` + ``--run-level full_model``. Example:
+    Nightly **Perf Test** jobs in ``test-nightly.yml`` use ``--test-config-file`` only (no ``-m``). E2e L4 function tests still use ``full_model`` + ``--run-level full_model`` (see [test-nightly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-nightly.yml)). Example:

 === "L5 level"

@@ -193,13 +186,9 @@ Failed jobs: 1/2

    ```

-    The latest L5 CI jobs are in [test-weekly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-weekly.yml):
-
-    - **Reliability** (+ invalid-parameter), **Perf Test** (Omni vLLM Text / Multi-Replica), and **Simple · CPU Coverage Test** → `WEEKLY=1` (or PR label `weekly-test` for Reliability/Perf). Scheduled upload of those env-gated steps still requires `main`.
-    - **E2E Tests** (slow Omni / TTS / Diffusion sweeps) → `NON_CRITICAL=1` (pipeline upload also accepts `WEEKLY=1` / `weekly-test`; the group `if` still needs `NON_CRITICAL`)
+    The latest L5 CI jobs (reliability + invalid-parameter weekly steps) are in [test-weekly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-weekly.yml).

 You can find more information about markers in [Test Writing Guide — Markers for Tests](./test_writing_guide.md#markers-for-tests).

 ## Adding New Test Cases
-
 Please refer to the [Test System Overview](./test_system_overview.md) and [Test Writing Guide](./test_writing_guide.md).
@@ -12,7 +12,8 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
      <th>Level</th>
      <th>Scope & Focus</th>
      <th>Model Coverage Strategy</th>
-      <th>Module Strategy</th>
+      <th>Feature Coverage Strategy</th>
+      <th>Interface Coverage Strategy</th>
      <th>Tags</th>
      <th>Time Cost</th>
      <th>Test Dir</th>
@@ -30,6 +31,7 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
      <td>/</td>
      <td>/</td>
      <td>/</td>
+      <td>/</td>
      <td><a href="https://github.com/vllm-project/vllm-omni/blob/main/.github/PULL_REQUEST_TEMPLATE.md">PR Checklist</a></td>
      <td>/</td>
      <td>/</td>
@@ -41,6 +43,7 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
      <td>/</td>
      <td>/</td>
      <td>/</td>
+      <td>/</td>
      <td><a href="./failures.md">CI Failures</a></td>
      <td>/</td>
      <td>/</td>
@@ -50,6 +53,7 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
      <td>Unit tests for components like entrypoints, models</td>
      <td>/</td>
      <td>/</td>
+      <td>/</td>
      <td><code>core_model and cpu</code></td>
      <td rowspan="2">&lt;15min</td>
      <td>
@@ -67,8 +71,9 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
    <tr>
      <td><strong>L2</strong><br>(E2E across models & GPU-required UT)</td>
      <td>Online (basic deployment scenarios):<br>dummy, normal inference function (output format, stream), some instance startup UT</td>
-      <td>Key models + online basic scenarios; request success, non-empty output, format match (no Whisper/accuracy)</td>
-      <td>Key features and modules that require launching instances (prefer random weights)</td>
+      <td>High-priority models + online basic scenarios; request success, non-empty output, format match (no Whisper/accuracy)</td>
+      <td>High-priority features and component tests that require launching instances</td>
+      <td>High-priority interfaces (using random lightweight models)</td>
      <td><code>core_model and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
      <td>
        <strong>Model E2E:</strong><br>
@@ -90,8 +95,9 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
    <tr>
      <td><strong>L3</strong><br>(Important Perf & Integration & Accuracy)</td>
      <td>Online & Offline (multiple deployment scenarios):<br>real model, normal inference function, normal accuracy</td>
-      <td>Key models + key online/offline scenarios; real weights, Whisper/similarity, preset voice gender, basic accuracy</td>
-      <td>Key features and modules that require launching instances (using real weights)</td>
+      <td>High/medium-priority models + key online/offline scenarios; real weights, Whisper/similarity, preset voice gender, basic accuracy</td>
+      <td>Medium-priority features and component tests that require launching instances</td>
+      <td>Medium-priority interfaces (using random lightweight models)</td>
      <td><code>advanced_model and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
      <td>&lt;30min</td>
      <td>
@@ -114,8 +120,9 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
    <tr>
      <td><strong>L4</strong><br>(Perf & Integration & Accuracy)</td>
      <td>Online: full functional scenarios + performance test + doc test + accuracy test</td>
-      <td>Key models: function, performance, accuracy, and doc testing</td>
-      <td>Other features and modules that require launching instances (using real weights)</td>
+      <td>High-priority models: function, performance, accuracy, and doc testing<br>Medium-priority models: function and doc testing</td>
+      <td>Low-priority features and component tests that require launching instances</td>
+      <td>Low-priority interfaces (using real weights)</td>
      <td><code>full_model and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
      <td>&lt;3 hour</td>
      <td>
@@ -140,10 +147,11 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
      <td>GPU</td>
    </tr>
    <tr>
-      <td><strong>L5</strong><br>(Stability & Reliability & selected Perf)</td>
-      <td>Online: long-term stability + reliability; coverage test</td>
-      <td>Long-term stability and reliability testing for key models<br>Non-critical scenario performance and accuracy testing for key models<br>Other models: function and doc testing</td>
+      <td><strong>L5</strong><br>(Stability & Reliability)</td>
+      <td>Online: long-term stability test + reliability test</td>
+      <td>Long-term stability and reliability testing for high-priority models<br>Low-priority models: function and doc testing</td>
      <td>/</td>
+      <td>Invalid-parameter validation for high-priority interfaces</td>
      <td><code>slow and hardware_test(H100, L4, etc.) and omni/tts/diffusion</code></td>
      <td> Depends on reality </td>
      <td>
@@ -170,8 +178,8 @@ For per-level test authoring (markers, examples), see [Test Writing Guide](./tes

 Before entering specific testing levels, the project establishes two common specifications aimed at standardizing the development process and quickly locating issues.

-1. ***PR Checklist ([`.github/PULL_REQUEST_TEMPLATE.md`](https://github.com/vllm-project/vllm-omni/blob/main/.github/PULL_REQUEST_TEMPLATE.md))***: This template defines the self-check items that must be completed before submitting a code review (Pull Request). It ensures that each code change meets basic requirements such as code style, dependency updates, and documentation synchronization before entering the automated testing pipeline, serving as the first manual line of defense for quality assurance.
-2. ***CI Failure Explanation ([CI Failures](./failures.md))***: This document archives and explains common failure patterns in the Continuous Integration (CI) pipeline, error log interpretation, and preliminary troubleshooting steps. It helps developers and testers quickly diagnose the causes of automated test failures, improving problem-solving efficiency.
+1.  ***PR Checklist ([`.github/PULL_REQUEST_TEMPLATE.md`](https://github.com/vllm-project/vllm-omni/blob/main/.github/PULL_REQUEST_TEMPLATE.md))***: This template defines the self-check items that must be completed before submitting a code review (Pull Request). It ensures that each code change meets basic requirements such as code style, dependency updates, and documentation synchronization before entering the automated testing pipeline, serving as the first manual line of defense for quality assurance.
+2.  ***CI Failure Explanation ([CI Failures](./failures.md))***: This document archives and explains common failure patterns in the Continuous Integration (CI) pipeline, error log interpretation, and preliminary troubleshooting steps. It helps developers and testers quickly diagnose the causes of automated test failures, improving problem-solving efficiency.

 ## Notes

@@ -509,7 +509,7 @@ L4 level testing is a comprehensive quality audit before a version release. It e
 #### 3.2 Testing Content and Scope

 - ***Full Functionality Testing***: Executes all test cases defined in `test_{model_name}_expansion.py`, covering all implemented features, positive flows, boundary conditions, and exception handling.
- ***Performance Testing***: JSON configs under `tests/dfx/perf/tests/` drive Omni, TTS, and diffusion serving benchmarks via `run_benchmark.py` or `run_diffusion_benchmark.py`, measuring throughput, latency, and related metrics under configured workloads.
+- ***Performance Testing***: Uses `tests/dfx/perf/tests/test_qwen3_omni_*.json` (Omni), `test_tts.json` / `test_voxcpm2.json` / `test_higgs_audio_v3.json` (TTS), and diffusion configs `tests/dfx/perf/tests/test_*_vllm_omni.json` (passed to `run_benchmark.py` or `run_diffusion_benchmark.py` via `--test-config-file` in nightly **Perf Test** steps) to drive throughput, latency, and memory benchmarks. Each JSON **case** may declare an optional top-level **`mark`** array: exactly one ``hardware_marks`` object plus pytest marker name strings (`full_model`, `omni` / `tts` / `diffusion`, …). Runners attach those marks to each parametrized `(server, benchmark index)` pair so **local** bulk runs can filter with `-m` (for example `-m "full_model and H100 and diffusion"`). Nightly CI perf jobs select workloads by **`--test-config-file`**, not `-m`. Details are in the Performance Tests example below.
 - ***Documentation Testing***: Verifies whether the example code provided to users is runnable and its results match the description.

 #### 3.3 Test Directory and Execution Files
@@ -567,7 +567,7 @@ L5 level testing focuses on the performance of model services under ***long-runn
 - ***Trigger Timing***: **`Weekly`** (weekly) or **`Days before Release`** (several days before a major release). Due to long execution times, the frequency is lower.
 - ***Run Command***:
    - ***Stability***: `pytest -s -v tests/dfx/stability/scripts/test_stability_qwen3_omni.py` or `pytest -s -v tests/dfx/stability/scripts/test_stability_wan22.py` (or add `test_stability_<model>.py` alongside a matching JSON config)
-    - ***Reliability***: `pytest -s -v tests/dfx/reliability/test_reliability_<model>.py -m slow` (current suites: `qwen3_omni`, `wan22`, `hunyuan_image`).
+    - ***Reliability***: `pytest -s -v tests/dfx/reliability/test_reliability_<model>.py -m slow` (current suites: `qwen3_omni`, `wan22`, `hunyuan_image`). Weekly CI (`.buildkite/cuda/test-weekly.yml`) runs one step per suite (`WEEKLY=1` or PR label `weekly-test`)
 - ***Script Example***:

 <details>
@@ -124,8 +124,8 @@ When `execute_model_batch` returns a `COMPUTE_DONE` with a single `async_output_
 1. **Drain Before Memory Release (`drain_async_outputs` / `_ASYNC_OUTPUT_DRAIN_TIMEOUT_S`)**:
   When workers transition into sleep states (`handle_sleep_task`) or invoke memory-releasing methods, `WorkerProc.drain_async_outputs()` blocks up to `_ASYNC_OUTPUT_DRAIN_TIMEOUT_S` (10.0s) until all in-flight background D2H/SHM packing tasks finish. This acts as a synchronization barrier preventing worker sleep cycles from releasing device tensors while the side CUDA stream is still actively reading them. Draining typically completes within 10–50 ms under normal operation.

-2. **Timeout Diagnostics (`_async_output_timeout()`)**:
-   The engine waits for `wait_output_ready()` with a timeout of `_async_output_timeout()` — 600.0s by default, overridable per run with the `VLLM_OMNI_ASYNC_OUTPUT_TIMEOUT` environment variable (a non-numeric or non-positive value is ignored with a warning and the default applies, since this resolves on the request path). If a timeout occurs, the executor logs detailed diagnostic bookkeeping (pending futures, cached outputs, and batch split maps) to identify whether the stall occurred in the worker, pump, or waiter. The packing itself takes milliseconds, but the wait is queued behind the step's GPU work, so its wall-clock duration tracks step time; the bound is generous because a hung engine is surfaced by the worker monitor and `check_health()`, not by this timeout.
+2. **Timeout Diagnostics (`_ASYNC_OUTPUT_TIMEOUT`)**:
+   The engine waits for `wait_output_ready()` with a timeout of `_ASYNC_OUTPUT_TIMEOUT` (30.0s). If a timeout occurs, the executor logs detailed diagnostic bookkeeping (pending futures, cached outputs, and batch split maps) to identify whether the stall occurred in the worker, pump, or waiter.

 3. **Queue Serialization & Atomic Future Resolution**:
   - `result_mq` writes are serialized with `_result_mq_lock` to protect the single-writer `MessageQueue` against concurrent writes from the worker main loop (`COMPUTE_DONE`) and background thread (`OUTPUT_READY`).
@@ -61,8 +61,7 @@ unstructured keyword dictionaries:
 - `quant` is consumed by FlashInfer and TRTLLM with backend-specific value
  validation;
 - `skip_softmax` is consumed by TRTLLM; and
- `block_sparse` is consumed by block-sparse backends such as RainFusion; and
- `fastvideo_vsa_topk` is consumed by FastVideo VSA.
+- `block_sparse` is consumed by block-sparse backends such as RainFusion.

 A backend reads only the fields it owns and rejects incompatible values. New
 options should be added to a shared typed spec only when more than one backend
@@ -88,19 +87,3 @@ An implementation change is complete only when it:
   corresponding user guide; and
 6. updates the overview matrix without moving algorithm details back into the
   landing page.
-
-## FastVideo VSA model metadata
-
-FastVideo VSA support is scoped to `FastVideo/FastWan2.2-TI2V-5B-Diffusers`, whose text-to-video and image-to-video modes both use `Wan22Pipeline`. It operates on a flattened DiT sequence but partitions tokens in
-the original latent video grid. Wan integrations therefore attach the
-post-patch `(T, H, W)` shape as `vsa_dit_seq_shape` attention metadata. The
-backend validates that `T * H * W` equals the sequence length, derives the
-runtime block count, and selects the configured top-k key/value blocks for
-each query block.
-
-Checkpoint-specific compensation remains model-owned. Wan layers pass an
-optional learned `gate_compress` projection through attention metadata when
-the checkpoint contains those weights. Checkpoints without the projection do
-not allocate or execute it. When top-k selects every block, native checkpoints
-route to SDPA, while FastVideo DMD checkpoints preserve the VSA all-block path
-to retain checkpoint semantics.
@@ -15,11 +15,8 @@ specified in the [Host Weight Runtime module design](../module/host_weight_runti
 ## Status

 The first implementation provides contracts and a CPU local-filesystem store.
-The initial diffusion consumer contract additionally defines typed,
-representation-independent final-layout identity/restoration mechanics plus a
-concrete BF16-with-preserved-FP32 policy for MiniMax H3. It remains
-library-only: no loader or DLO path selects, publishes, restores, or transports
-that artifact yet.
+It does not enable cached loading for a model by itself. A consumer must still
+provide an exact identity, a representation producer, and a model restorer.

 V1 includes:

@@ -33,9 +30,8 @@ V1 includes:

 V1 does not include:

- a public CLI, loader activation, or default-on model integration;
- lease handoff to DLO or another transport consumer;
- FP8, quantized, merged-adaptation, or additional model producers;
+- a public CLI or default-on model integration;
+- a generic BF16, FP8, quantized, or model-specific producer;
 - CUDA registration, pinned staging, H2D scheduling, or GPU kernels;
 - a remote artifact provider or cross-node coordination;
 - automatic eviction; or
@@ -220,56 +216,6 @@ Dynamic LoRA overlays are not part of a reusable base-weight artifact. A
 statically merged adapter is cacheable only as a separate identity containing
 the adapter fingerprint and merge semantics.

-### Initial diffusion final-layout contract
-
-The shared diffusion contract covers complete final-layout DiT parameters and
-persistent buffers. Text encoders, VAEs, non-persistent derived state, and other
-pipeline components remain outside the artifact. One explicit representation
-policy selects allowed dtypes, tensor roles, physical layout identity, producer
-ABI, manifest schema, and restoration schema.
-
-The contract is intentionally separate from loader activation:
-
- `FinalLayoutRequest` contains typed loader identity/configuration fingerprints,
-  TP coordinate, and conservative SP semantics. It has no open metadata bag,
-  DP coordinate, SP rank, device identity, DLO transfer mode, registration
-  policy, or store path.
- `FinalLayoutArtifactSpec` binds one `WeightRepresentation` and runtime-layout
-  name to explicit producer/restorer schemas and a canonical, versioned
-  implementation ABI descriptor. Compatibility never depends on reflective
-  source inspection.
- `PreparedWeightSource` snapshots immutable revisions or exact local file
-  content plus a typed checkpoint-adapter identity before ordinary
-  materialization. Source replacement before or during production fails
-  publication. A hash-looking symlink basename is trusted only for an explicit
-  Hugging Face Hub source whose repository ID and
-  `models--.../snapshots/<revision> -> blobs/<hash>` topology validate; every
-  local or otherwise unverified symlink target is content-hashed.
- the tensor ownership digest records exact runtime names, kinds, shapes,
-  semantic roles, dtypes, and strides from a CPU or meta model skeleton;
- `FinalLayoutTensorRestorer` accepts only an exact lease identity, validates
-  complete policy-defined coverage without mutation, and returns a one-shot
-  commit plan;
- each model declares one dtype-neutral `FinalLayoutModelContract` with an
-  explicit implementation version and a post-commit validator; and
- `FinalLayoutBF16Producer` accepts only the matching identity context and a
-  finalized CPU model. It is `POST_LOAD_ONLY` and `SINGLE_PROCESS` per exact TP
-  coordinate. Its BF16 policy preserves model-declared FP32 parameters and
-  buffers and revalidates MiniMax H3 mixed-precision invariants.
-
-Other representations reuse source identity, typed parallel identity, tensor
-ownership, and exact restoration only when their policy proves those semantics.
-For example, runtime FP8 needs a separate policy/producer for generated scales,
-quantization metadata, and Cutlass physical layouts; it is not enabled by
-changing a dtype string on the BF16 producer.
-
-This stage makes no startup, sharing, or DLO performance claim. A following
-consumer PR owns disabled/preferred/required precedence, mixed-component loader
-transactions, warm-hit restoration, and transactional lease handoff. A TP2
-prewarm deployment will require a matching TP2 producer cohort to populate both
-TP-coordinate identities even though the store coordinates each artifact
-independently.
-
 ## Host sharing and GPU transport

 Every process receives its own virtual mappings and `HostWeightLease`, but
@@ -20,7 +20,7 @@ implemented but has less end-to-end coverage than the primary path.
 Legend: ✅ supported, ⚠️ compatibility path or limited validation, ❌ unsupported.

 | Feature | DLO + AllGather | DLO without AllGather |
-| --- | --- | --- |
+|---|---|---|
 | **DP** | ✅ Primary path; host weights are sharded across the DP group. | ✅ Each DP rank streams complete rank-local blocks. |
 | **SP** | ✅ When DP=1, DLO uses the SP group for weight sharding. | ✅ SP remains active without a DLO weight collective. |
 | **TP > 1** | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. |
@@ -169,192 +169,10 @@ This mode means:
  copy per rank.
 - The scheduler does not require a synchronized DP request wave for DLO.

-### Component allocator-cache retention
-
-MiniMax-H3 DLO keeps one bounded PyTorch allocator cache across its staged
-encoder, DiT, and VAE component boundaries so allocations released by one
-component can be reused by the next. The DiT prefetch path and its two shared
-device buffers are unchanged.
-
-After a component is offloaded, cached-but-unallocated memory is retained only
-while it is at most 25% of device capacity and at least 5% of device capacity
-remains physically free. Crossing either bound releases the allocator cache.
-Missing allocator telemetry also releases it conservatively. Component or
-staging failure forces a release, and an out-of-memory allocation gets one
-retry after release. This is a component-local policy rather than a global
-`empty_cache` override, so the executor's unconditional shutdown cleanup is
-preserved.
-
-Retaining allocator blocks trades higher reserved and peak device memory for
-lower lifecycle latency. The benefit is device- and workload-dependent, and a
-smaller or externally loaded device may cross a bound and return to the
-ordinary per-component cache-release behavior. Current performance validation
-covers MiniMax-H3 no-AllGather DLO on CUDA. Other MiniMax-H3 DLO transfer
-topologies and platforms receive the same bounded policy but are not claimed
-performance targets; other DLO models remain unchanged unless they explicitly
-adopt it.
-
-### Final-layout Host Weight Runtime consumer
-
-> **Status:** PR2 implementation. The representation-independent identity,
-> producer, and restoration contracts landed in
-> [PR #6445](https://github.com/vllm-project/vllm-omni/pull/6445); this consumer
-> wires the final-layout BF16 path into eligible no-AllGather DLO.
-
-Final-layout Host Weight Runtime (HWR) backing is opt-in and currently applies
-only to no-AllGather DLO. Enable it with
-`--host-weight-runtime-mode preferred` (or `required`) and
-`--host-weight-runtime-root <node-local-root>`. Registration of shared mappings
-and direct asynchronous H2D remain a separate PR3 transport change.
-
-The modes express operator fallback policy, not different artifact formats:
-
- `preferred` consumes an exact hit, but on a miss it allows canonical loading
-  followed by post-load publication. It is the normal population path.
- `required` consumes the same exact artifact but fails startup on a miss or
-  unusable artifact. It never invokes canonical DiT fallback or post-load
-  publication and therefore cannot populate an empty store.
-
-PR2 has no separate prewarm command. Operators populate one matching producer
-cohort per node-local storage domain in `preferred` mode, then restart the same
-model revision and parallel layout with `required`. TP coordinates have distinct
-identities, while equivalent DP replicas share them.
-
-The current producer/consumer boundary is the model-declared final-layout BF16
-contract (MiniMax H3 today). It supports ordinary-loader final layouts for TP1
-and TP2 rank identities plus SP layout identities; online quantization, HSDP,
-LoRA/adapted weights, and non-default load formats remain ineligible. HWR mode
-`disabled`, DLO-disabled, and DLO AllGather configurations stop before source
-identity or store construction and retain the existing checkpoint-mmap or
-ordinary-loader path.
-
-Eligibility is decided before constructing HWR, resolving canonical sources,
-hashing identity inputs, probing a filesystem, or emitting an HWR observer
-event:
-
-```text
-DLO disabled / HWR disabled / DLO AllGather
-  -> zero final-layout HWR interaction
-  -> preserve current checkpoint-mmap or ordinary-loader behavior
-```
-
-When final-layout HWR is explicitly enabled for no-AllGather DLO, runtime
-outcomes are authoritative:
-
- `LOCAL_HIT`: plan and commit an exact restoration, then transfer the lease
-  transactionally to DLO.
- `CANONICAL_FALLBACK`: bypass checkpoint mmap, canonically materialize and
-  finalize the DiT, then call `publish_after_load()` for a future startup. The
-  current startup retains its canonical tensors.
- `FAILED`: fail startup. Preferred mode does not reinterpret nonretryable
-  identity, configuration, or compatibility failures as misses.
- Required mode cannot bootstrap an empty store through a `POST_LOAD_ONLY`
-  producer.
-
-Checkpoint mmap remains an unchanged control path whenever final-layout HWR is
-not selected.
-
-#### Pre-service transaction boundary
-
-The startup transaction does not end at restore commit. A lease-backed model is
-disposable until backend setup and initial prefetch complete:
-
-```text
-UNRESOLVED
-  -> LEASE_OWNED_BY_LOADER
-  -> RESTORE_PLANNED
-  -> COMMIT_STARTED
-  -> CARRIER_OWNED
-  -> BACKEND_OWNED
-  -> BACKEND_READY
-  -> IN_SERVICE
-```
-
-Failure before `COMMIT_STARTED` closes the loader-owned lease. Preferred mode
-may canonically load using the untouched skeleton; required mode fails.
-
-Failure from `COMMIT_STARTED` through `BACKEND_READY` must:
-
-1. synchronize or otherwise quiesce partial backend and initial-prefetch work;
-2. release hook, staging, restore-plan, and model references;
-3. close the lease through its current owner;
-4. discard the restored model; and
-5. in preferred mode, construct a fresh canonical model while bypassing HWR
-   lookup, HWR publication, and checkpoint mmap for that recovery attempt.
-
-Required mode fails instead of constructing the fresh fallback. Once the model
-enters service, failures follow normal runtime handling rather than startup
-fallback.
-
-#### Finalization phases
-
-Cold canonical loading separates byte-changing work from shared runtime-state
-finalization:
-
-```text
-Cold-only byte-changing:
-  casting, reordering, packing, quantization, generated scales,
-  model-specific weight transforms, and calibration that mutates
-  parameters or persistent buffers
-
-Shared cold/warm non-byte:
-  validation, hook installation, eval state, bookkeeping,
-  and non-persistent runtime state
-```
-
-The warm path must snapshot restored tensor bytes and backing pointers before
-shared finalization and prove both are unchanged afterward.
-
-#### Lease ownership
-
-Lease transfer is single-owner and single-take:
-
-```text
-loader owns
-  -> restore plan borrows
-  -> commit succeeds
-  -> runner carrier takes once
-  -> backend takes once before asynchronous work
-  -> backend drains pending H2D work
-  -> backend releases hooks, staging, and model references
-  -> backend closes lease
-```
-
-The runner carrier is process-local, rejects serialization, and rejects a
-duplicate `take()`. If the carrier still owns the lease, runner cleanup closes
-it. Once the backend takes ownership, only backend abort or teardown may drain
-work and close it.
-
-The implementation keeps the final-layout lease through backend setup and
-initial prefetch. The backend uses the existing two bounded rank-local staging
-slots for the immutable HWR tensors, and its transactional `enable()` cleanup
-removes partial hooks and closes the lease before reporting a setup failure.
-Preferred mode then uses the runner's fresh canonical retry; required mode
-propagates the failure.
-
-#### PR2 promotion gates
-
- Warm hit performs zero ordinary DiT materialization and zero producer calls.
- Shared warm finalization changes neither restored bytes nor backing pointers.
- Preferred `FAILED` outcomes do not silently fall back.
- Planning failure leaves the skeleton untouched.
- Any pre-service failure after commit begins discards the restored model;
-  preferred mode uses a fresh canonical model and required mode fails.
- Mixed components load normally and no required tensor remains on `meta`.
- Disabled mode and AllGather emit zero final-layout HWR interaction.
- The lease carrier rejects duplicate take and serialization.
- Backend setup or prefetch failure drains asynchronous work before lease close.
- Compatible checkpoint mmap remains an unchanged benchmark and control path.
- Prewarm uses one matching producer cohort per TP/SP topology and storage
-  domain.
- Benchmarks compare ordinary loading, cold publication, and warm staging with
-  output parity, startup latency, aggregate PSS, page-cache state, HBM, and H2D
-  payload.
-
 ## Parallelism compatibility

 | Parallelism | DLO + AllGather | DLO without AllGather |
-| --- | --- | --- |
+|---|---|---|
 | **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages on each node; fallback runtime tensors remain private. |
 | **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active, but DLO keeps standard-loader rank-local weights and adds no SP weight collective. |
 | **TP > 1** | Outside the Phase A shared-mmap support scope. The loader falls back before mutation, preserves TP-local layouts, and DLO may apply DP/SP host sharding to those ordinary runtime tensors. | Outside the Phase A shared-mmap support scope. The ordinary TP-aware loader produces rank-local tensors, which DLO streams without an additional weight collective; DP replicas retain private runtime storage. |
@@ -367,8 +185,7 @@ propagates the failure.
  the SP group becomes DLO's sharding group in AllGather mode.
 - **DP + TP/SP without AllGather:** standard model loading defines the
  rank-local tensor layout. DLO adds no cross-DP, cross-TP, or cross-SP weight
-  collective. When the model declares the final-layout contract, HWR keys the
-  reusable artifact by TP rank/size and SP layout.
+  collective.
 - **HSDP + SP:** the general parallel configuration permits HSDP over SP, but
  DLO must use `--dlo-no-use-allgather`. HSDP remains responsible for weight
  materialization and synchronization.
@@ -394,11 +211,10 @@ reconstructs those tensors with their recorded layouts. Other online methods
 must use `--dlo-no-use-allgather` or disable online quantization until their
 runtime layouts are validated.

-The Host Weight Runtime representation and publication contracts are merged;
-see [RFC #6414](https://github.com/vllm-project/vllm-omni/issues/6414) and
-[PR #6445](https://github.com/vllm-project/vllm-omni/pull/6445). The planned
-no-AllGather consumer above is the PR2 implementation; registration and direct
-asynchronous H2D remain a separate transport follow-up.
+A normalized runtime mmap cache, built through the ordinary loader, is the
+proposed general mechanism for sharing transformed TP or quantized layouts.
+That cache and its publication/lifecycle protocol are intentionally outside
+this phase; see [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).

 ## Validation coverage

@@ -415,8 +231,6 @@ Current source-level validation includes:
  transforms without parameter-side flags;
 - resident-layer requests requiring no-AllGather;
 - DP request-wave validation for denoising-step compatibility;
- bounded component allocator-cache retention, conservative/forced release,
-  OOM retry, and immutable encoder non-block staging;
 - sharding, double-buffer, AllGather-size, and heterogeneous-block regression
  tests.

@@ -429,7 +243,7 @@ VAEs at TP1. They validate the ordinary-loader fallback only, not direct mmap
 or shared-mmap host-memory savings.

 | Configuration | Result | Warm E2E | Peak device memory | Host PSS |
-| --- | ---: | ---: | ---: | ---: |
+|---|---:|---:|---:|---:|
 | DP4xTP1 AllGather | Passed, 4 concurrent requests | 2.87 s / 4 requests | 13.84 GiB | 211.99 GiB |
 | DP4xTP1 no-AllGather | Passed, 1 request | 15.02 s | 13.23 GiB | 187.77 GiB |
 | DP2xTP2 AllGather | Passed, 2 concurrent requests | 4.16 s / 2 requests | 12.50 GiB | 211.97 GiB |
@@ -461,7 +275,7 @@ pipeline components, so each worker should be compared with the same worker in
 the other storage mode.

 | Worker | Ordinary RSS | mmap RSS | Ordinary PSS | mmap PSS | PSS reduction |
-| --- | ---: | ---: | ---: | ---: | ---: |
+|---|---:|---:|---:|---:|---:|
 | DP worker 0 | 168.27 GiB | 132.76 GiB | 167.84 GiB | 101.43 GiB | 66.40 GiB |
 | DP worker 1 | 116.19 GiB | 79.97 GiB | 115.73 GiB | 48.64 GiB | 67.09 GiB |
 | **Two-worker total** | — | — | **283.56 GiB** | **150.08 GiB** | **133.48 GiB (47.1%)** |
@@ -18,7 +18,7 @@ validation_paths:
  - tests/diffusion/offloader/**
 upstream_refs:
  - torch.nn.Module.to
-last_reviewed: 2026-08-25
+last_reviewed: 2026-07-16
 ---

 # Diffusion offloader
@@ -48,11 +48,6 @@ and correctness required by the selected execution mode.
 **Rule:** Retained host and device copies MUST respect configured limits and
 provide deterministic teardown.

-Allocator-cache retention MUST remain local to an explicit component owner,
-declare both cache and physical-free-memory bounds, release conservatively
-when telemetry is unavailable, and force release on failure or memory
-pressure. It MUST NOT replace unconditional executor shutdown cleanup.
-
 ## Safe-change guide

 Test enabled and disabled paths, repeated execution, asynchronous transfer,
@@ -56,8 +56,7 @@ The contracts have deliberately narrow ownership:

 - `HostWeightRuntime` applies resolution and fallback policy and emits one
  terminal resolution report. Explicit post-load publication emits a separate
-  report and closes its store-returned lease before returning; it does not
-  revise the completed resolution.
+  report and may return a lease, but does not revise the completed resolution.
  The runtime does not own a model or call the canonical loader.
 - `HostWeightStore` owns artifact lookup, coordinated construction,
  publication, validation, quarantine, and lifecycle.
@@ -52,8 +52,6 @@ th {
 | `LTX2DistilledOneStagePipeline` | LTX-2.5 merged-distilled one-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
 | `LTX2DistilledTwoStagePipeline` | LTX-2.5 merged-distilled two-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
 | `LingBotVideoPipeline` | LingBot-Video dense and MoE T2I, T2V, TI2V | `robbyant/lingbot-video-dense-1.3b`, `robbyant/lingbot-video-moe-30b-a3b` | ✅︎ | | | | — |
-| `SanaVideoPipeline` | SANA-Video-2B T2V (native 480p/720p; Diffusers adapter validated at both resolutions) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers`, `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-Video-2B.md) |
-| `SanaImageToVideoPipeline` | SANA-Video-2B I2V (native and Diffusers adapter validated at 480p/720p) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers`, `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-Video-2B.md) |
 | `MiniMaxH3Pipeline` | MiniMax H3 T2VA, first/last-keyframe FL2VA, and mixed-reference Ref2VA | `MiniMaxAI/MiniMax-H3` | ✅︎ | ✅︎ | | | [Published](https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3) |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Pi0Pipeline` | π0 (Pi-Zero) VLA | `lerobot/pi0_base` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/lerobot/Pi0.md) |
@@ -70,12 +68,17 @@ th {
 | `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | | — |
 | `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | | — |
 | `DotsTTSForConditionalGeneration` | dots.tts | `rednote-hilab/dots.tts-soar` | ✅︎ | | | | — |
+| `MammothModa2ForConditionalGeneration` | MammothModa2-Preview | `bytedance-research/MammothModa2-Preview` | ✅︎ | ✅︎ | | | — |
+| `MammothModa2ForConditionalGeneration` | MammothModa2-Dev (AR-only image understanding) | `bytedance-research/MammothModa2-Dev` | ✅︎ | | | | — |
 | `Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `FluxKontextPipeline` | FLUX.1-Kontext-dev | `black-forest-labs/FLUX.1-Kontext-dev` | ✅︎ | ✅︎ | | | — |
 | `FluxPipeline` | FLUX.1-dev | `black-forest-labs/FLUX.1-dev` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `FluxPipeline` | FLUX.1-schnell | `black-forest-labs/FLUX.1-schnell` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `OmniGen2Pipeline` | OmniGen2 | `OmniGen2/OmniGen2` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `StableAudioPipeline` | Stable-Audio-Open | `stabilityai/stable-audio-open-1.0` | ✅︎ | ✅︎ | | ✅︎ | — |
+| `SoulXSingerPipeline` | SoulX-Singer (SVS) | `Soul-AILab/SoulX-Singer` | ✅︎ | | | | — |
+| `SoulXSingerSVCPipeline` | SoulX-Singer-SVC | `Soul-AILab/SoulX-Singer` (`model-svc.pt`) | ✅︎ | | | | — |
+| `AudioXPipeline` | AudioX | `zhangj1an/AudioX` | ✅︎ | ✅︎ | | | — |
 | `MiniMaxMusic3ForConditionalGeneration` | MiniMax Music 3 (text-to-music) | `MiniMaxAI/MiniMax-Music3` | ✅︎ | | | | — |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` | ✅︎ | | | | [Published](https://recipes.vllm.ai/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice) |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -106,7 +109,6 @@ th {
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `GepardTalkerForConditionalGeneration` | Gepard-1.0 | `nineninesix/gepard-1.0` | ✅︎ | | | | — |
 |`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | | — |
-| `HiDreamO1ImagePipeline` | HiDream-O1-Image | `HiDream-ai/HiDream-O1-Image` | ✅︎ | | | | — |

 The `Recipe` column links the corresponding deployment recipe. `Published` is
 preferred when a corresponding `recipes.vllm.ai` page exists; otherwise the
@@ -131,7 +131,7 @@ Content-Type: application/json

 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `ref_audio` | string | null | Reference audio (HTTP URL, base64 data URL, or `file://` URI with `--allowed-local-media-path`). Local files fold `mtime_ns` and `size` into cache keys to automatically reload on-disk edits; HTTP URLs and base64 URIs remain cached by string locator. |
+| `ref_audio` | string | null | Reference audio (HTTP URL, base64 data URL, or `file://` URI with `--allowed-local-media-path`) |
 | `ref_text` | string | null | Transcript of reference audio |
 | `x_vector_only_mode` | bool | null | Use speaker embedding only (no ICL) |

@@ -20,7 +20,6 @@ For the internal selector, registry, and platform contract, see
 | Install and use SageAttention 2.2 or SageAttention3 | [SageAttention](attention_backends/sage.md) |
 | Match training or rollout kernels loaded from Hugging Face | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
 | Use block-sparse video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |
-| Use FastVideo VSA with FastWan2.2-TI2V-5B on CUDA | [FastVideo VSA](attention_backends/fastvideo_vsa.md) |

 ## Backend options

@@ -36,7 +35,6 @@ For the internal selector, registry, and platform contract, see
 | `FLASH_ATTN_HUB` | Hub kernel | FlashAttention 2 from Hugging Face `kernels` | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
 | `FLASH_ATTN_3_HUB` | Hub kernel | FlashAttention 3 from Hugging Face `kernels` on Hopper or newer | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
 | `RAINFUSION_ATTN` | Block sparse | MindIE-SD RainFusion video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |
-| `FASTVIDEO_VSA` | Block sparse | FastVideo variable sparse self-attention for FastWan2.2-TI2V-5B on CUDA | [FastVideo VSA](attention_backends/fastvideo_vsa.md) |

 ## Configuration

@@ -107,7 +105,6 @@ Backend-specific typed blocks are documented with their consumers:
  and [TRTLLM SAGE](attention_backends/trtllm.md#sage-quantization).
 - `skip_softmax`: [TRTLLM Skip-Softmax](attention_backends/trtllm.md#skip-softmax).
 - `block_sparse`: [RainFusion](attention_backends/rainfusion.md#configuration).
- `fastvideo_vsa_topk`: [FastVideo VSA](attention_backends/fastvideo_vsa.md#choose-top-k).

 ## Platform defaults

@@ -1,125 +0,0 @@
-# FastVideo VSA
-
-FastVideo Variable Sparse Attention (VSA) accelerates the `FastVideo/FastWan2.2-TI2V-5B-Diffusers` model's
-self-attention by partitioning the post-patch latent grid into spatiotemporal
-blocks. For every query block, VSA scores the key/value blocks and computes
-attention only against the selected top-k blocks.
-
-The supported checkpoint provides both text-to-video and image-to-video modes through `Wan22Pipeline`; the separate Wan I2V-14B, S2V, and VACE pipelines are outside this backend's supported scope.
-
-VSA is a CUDA-only, explicitly selected backend. It requires the
-`fastvideo-kernel` package and currently supports non-causal self-attention
-with equal query and key/value sequence lengths. Unsupported shapes, masks,
-dtypes, sequence-parallel execution, or kernel failures fall back to
-`TORCH_SDPA` and emit a warning with the reason.
-
-## Enable the backend
-
-For online serving, select the backend with the existing attention backend
-flag. Use `--fastvideo-vsa-topk` to set the number of key/value blocks retained
-for every query block:
-
-```bash
-vllm-omni serve <model> \
-  --diffusion-attention-backend FASTVIDEO_VSA \
-  --fastvideo-vsa-topk 64
-```
-
-The backwards-compatible environment variable selects the backend with the
-default `topk=64`:
-
-```bash
-export DIFFUSION_ATTENTION_BACKEND=FASTVIDEO_VSA
-vllm-omni serve <model>
-```
-
-To tune top-k, pass the CLI backend and top-k flags together as shown above,
-or use the equivalent structured configuration:
-
-```bash
-vllm-omni serve <model> \
-  --diffusion-attention-config \
-  '{"default":{"backend":"FASTVIDEO_VSA","fastvideo_vsa_topk":64}}'
-```
-
-Do not combine `--diffusion-attention-backend` with an explicit
-`diffusion_attention_config.default.backend`. The top-k value must be positive
-and is valid only when the default backend is `FASTVIDEO_VSA`.
-
-For a deploy YAML stage, use either the shorthand fields:
-
-```yaml
-stages:
-  - stage_id: 0
-    diffusion_attention_backend: FASTVIDEO_VSA
-    fastvideo_vsa_topk: 64
-```
-
-or the structured configuration:
-
-```yaml
-stages:
-  - stage_id: 0
-    diffusion_attention_config:
-      default:
-        backend: FASTVIDEO_VSA
-        fastvideo_vsa_topk: 64
-```
-
-## Choose top-k
-
-At runtime the backend logs the sequence shape and derived block count:
-
-```text
-FASTVIDEO_VSA routing: seq_len=27280, dit_seq_shape=(31, 22, 40),
-block_size=(4, 8, 8), num_blocks=120, topk=64, keep_ratio=53.3%,
-checkpoint_mode=native, route=VSA
-```
-
-Use `num_blocks` as the upper bound when tuning top-k. Top-k is the number of
-key/value blocks retained for each query block, not the number of tokens or
-the total number of blocks processed by the layer.
-
- A smaller top-k increases sparsity and may reduce attention computation, but
-  it can remove relevant blocks and reduce visual quality or temporal
-  consistency. Routing and padding overhead can also make a smaller value
-  slower for some shapes.
- A larger top-k retains more context and generally approaches dense-attention
-  quality, but reduces the potential speedup and increases memory traffic.
- `topk > num_blocks` is invalid for the runtime shape and falls back to SDPA.
- For a native checkpoint, `topk == num_blocks` routes to SDPA because scoring
-  every block provides no sparsity benefit.
- For a FastVideo DMD checkpoint, `topk == num_blocks` stays on the VSA
-  all-block path so the checkpoint keeps its trained compensation semantics.
-
-There is no universally optimal value. Resolution, frame count, GPU, kernel
-version, and checkpoint all affect both quality and latency. Start from the
-logged `num_blocks`, test several keep ratios on the target workload, and
-compare output quality against the same checkpoint running dense attention.
-
-## Checkpoint behavior
-
-The backend does not expose a user-selectable gate mode. Wan checkpoints that
-contain learned `to_gate_compress` weights use that projection automatically.
-When those weights are absent, the unused projection is removed and VSA uses
-the sparse branch without learned compensation.
-
-FastVideo DMD checkpoints use their fixed distilled timestep schedule. Native
-Wan checkpoints keep their normal scheduler and inference-step configuration;
-selecting VSA does not turn a native checkpoint into a distilled model.
-
-## Verify routing and fallback
-
-Check the startup and first-forward logs instead of assuming that selecting
-the backend guarantees sparse execution:
-
- `route=VSA` means top-k block selection is active.
- `route=VSA_ALL_BLOCKS` means the FastVideo DMD checkpoint retained all
-  blocks through the VSA kernel.
- `route=SDPA` or `FASTVIDEO_VSA falling back to SDPA: ...` means dense SDPA
-  executed; the warning includes the reason.
-
-The current kernel requires CUDA tensors in FP16 or BF16, 256-token blocks,
-standard `head_size**-0.5` scaling, equal Q/K/V head counts, no attention mask,
-and no active sequence-parallel context. NPU and XPU paths do not execute the
-FastVideo VSA CUDA kernel.
@@ -49,6 +49,7 @@ higher-priority strategy is selected.
 | StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | — | yes | — | `transformer_blocks` |
 | Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | yes | yes | — | `blocks` |
 | Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | yes | yes | — | `blocks` |
+| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` | yes | yes | — | `layers` |
 | BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | — | yes | — | `layers`, customized modules |
 | Cosmos3OmniDiffusersPipeline | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | `Cosmos3VFMTransformer`, `Cosmos3LanguageModel` | yes | yes | yes | `layers`, `gen_layers` |

@@ -101,16 +101,8 @@ set `DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA` or configure
 `--max-num-seqs >1`. See the
 [HunyuanImage-3.0 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Tencent/HunyuanImage-3.0-Instruct.md)
 for its validated configuration. Helios supports single-request step execution only: use
-`--step-execution --max-num-seqs 1` for Helios. MiniMax H3 supports step-wise
-continuous batching by packing co-batched requests into one sequence that keeps
-a separate attention document per request; that layout needs a backend which
-honors the packed `cu_seqlens` metadata, so run it with
-`--diffusion-attention-backend FLASH_ATTN` (other backends stay correct but
-fall back to one transformer forward per request). Batching H3 does not improve
-its throughput — see the measured numbers in the
-[MiniMax-H3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/MiniMaxAI/MiniMax-H3.md)
-— so keep `--max-num-seqs 1` unless you need step-level scheduling. Consult the
-selected pipeline's documentation and source for the latest support status.
+`--step-execution --max-num-seqs 1` for Helios. Consult the selected pipeline's
+documentation and source for the latest support status.

 ## Streaming Output

@@ -63,8 +63,6 @@ omni = Omni(
 | `--dlo-use-allgather` | Shard host weights and reconstruct with AllGather | `true` |
 | `--dlo-no-use-allgather` | Stream complete rank-local blocks without a DLO weight collective | `false` |
 | `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |
-| `--host-weight-runtime-mode {disabled,preferred,required}` | HWR policy: no interaction, populate on a miss, or require an exact hit | `disabled` |
-| `--host-weight-runtime-root PATH` | Writable node-local HWR store shared by workers in one storage domain; required for `preferred` and `required` | unset |

 ## Host-weight loading

@@ -106,88 +104,6 @@ process.
 When the effective DLO group size is one, `dlo_use_allgather=True` does not
 perform a collective and uses the same rank-local transfer behavior.

-### Final-layout Host Weight Runtime
-
-HWR is an opt-in startup optimization for models that declare the final-layout
-BF16 restore contract. Use it only with no-AllGather DLO:
-
-```bash
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --dlo-no-use-allgather \
-  --host-weight-runtime-mode preferred \
-  --host-weight-runtime-root /var/cache/vllm-omni/hwr
-```
-
-#### Choosing a mode
-
-| Mode | Exact local artifact hit | Miss or recoverable artifact/store problem | Intended use |
-| --- | --- | --- | --- |
-| `disabled` | HWR is not consulted | Use the existing checkpoint-mmap or ordinary-loader path | Default compatibility path |
-| `preferred` | Restore the final-layout artifact | Load canonically, serve with those tensors, and attempt to publish an artifact for the next startup | Normal deployment and store population |
-| `required` | Restore the final-layout artifact | Fail startup without canonical DiT fallback or publication | Enforce a pre-populated store in controlled rollouts or CI |
-
-Both enabled modes still fail on non-retryable configuration, identity, or
-compatibility errors. `preferred` is a fallback policy for a cache miss or a
-recoverable cache problem; it does not hide an invalid deployment.
-
-#### Populating a store for `required` mode
-
-`required` is deliberately consume-only. PR2 does not include a separate
-prewarm command, so populate each node-local storage domain with one matching
-`preferred` producer cohort:
-
-1. Choose a persistent, writable root visible to every diffusion worker in the
-   storage domain. Do not use a process-private temporary directory.
-2. Start the deployment with `preferred`, using the exact model revision,
-   dtype, TP size, SP configuration, and other weight-layout settings intended
-   for serving. No inference request is needed; publication happens during
-   model startup. Wait for the engine to become healthy, then shut it down
-   cleanly.
-3. Restart with the same arguments and root, changing only the mode to
-   `required`. A successful startup proves that every worker acquired a valid
-   artifact; a miss, corrupt artifact, or incompatible identity fails startup.
-4. Repeat the population start on every node or storage domain because this
-   store is node-local.
-
-For example:
-
-```bash
-# First startup: canonically load and populate the exact artifacts.
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --dlo-no-use-allgather \
-  --host-weight-runtime-mode preferred \
-  --host-weight-runtime-root /var/cache/vllm-omni/hwr
-
-# After a healthy startup and clean shutdown, enforce cache hits.
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --dlo-no-use-allgather \
-  --host-weight-runtime-mode required \
-  --host-weight-runtime-root /var/cache/vllm-omni/hwr
-```
-
-Include the same TP and SP layout flags in both commands. DP rank and DP size
-are excluded from artifact identity, so the population and serving DP sizes may
-differ and equivalent DP replicas share artifacts. TP rank is included, so a
-TP-N deployment normally needs N rank-specific artifacts in each storage
-domain; launching the matching TP cohort creates that set. If the model
-revision or layout changes, run `preferred` again for the new identity before
-returning to `required`.
-
-On a cold start, the canonical loader remains authoritative and publishes a
-validated final-layout artifact for later workers. A warm start restores the
-DiT final tensors without ordinary DiT materialization, then DLO streams them
-through the same two bounded host staging slots. The artifact identity includes
-the TP rank/size and SP layout, so TP1, TP2 rank-local shards, and distinct SP
-layouts do not alias one another. Publication failure remains separate from a
-`preferred` serving startup; the next `required` startup provides the explicit
-artifact-availability check.
-
-When HWR is disabled, DLO is disabled, or AllGather is enabled, the loader does
-not resolve HWR sources or construct its store.
-
 ## Declarative topology

 Models may declare an `OffloadPlan` instead of embedding offload logic:
@@ -219,9 +135,7 @@ must enter each collective.
 - Direct checkpoint mmap currently requires TP1. TP greater than one is
  outside the Phase A shared-mmap support scope and falls back before model
  mutation to the ordinary TP-aware loader. DLO can stream that runtime layout,
-  and eligible no-AllGather configurations may use HWR final-layout artifacts,
-  but direct checkpoint mmap provides no shared-mmap host-memory guarantee for
-  TP greater than one.
+  but it provides no shared-mmap host-memory benefit or guarantee.
 - HSDP plus AllGather is rejected to avoid double sharding. HSDP without
  AllGather has limited end-to-end validation.
 - Per-tensor online FP8 linears use the ordinary loader and can run with either
@@ -233,10 +147,10 @@ must enter each collective.
  `OffloadPlan` that declares eligible `resident_dit_paths`.
 - DP concurrency requires an explicit, identical inference-step count.

-Sharing quantized or otherwise unvalidated transformed layouts through a
-normalized HWR producer is a follow-up design in
+Sharing transformed TP or quantized runtime layouts through a normalized mmap
+cache is a follow-up design in
 [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195), not part of
-the current BF16 final-layout path.
+the direct-checkpoint path.

 See the [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
 for an end-to-end example.
@@ -124,12 +124,12 @@ The following tables show which models support each feature:
 | **FLUX.2-dev**           |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |          ✅           |       ❌        |        ❌         |
 | **GLM-Image**            |     ❌     |     ❌      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Hidream-I1-Full**        |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
-| **HiDream-O1-Image**     |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **HunyuanImage3**        |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ✅        |        ✅*        |
 | **Krea 2**               |     ❌     |     ✅      |           ❌           |       ❌        |         ❌         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LongCat-Image**        |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **LongCat-Image-Edit**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **MagiHuman**            |     ❌     |     ❌      |           ❌           |       ❓        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **MammothModa2(T2I)**    |     ❌     |     ❌      |           ❌           |       ❌        |         ❌         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Nextstep_1(T2I)**      |     ❓     |     ❓      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **OmniGen2**             |     ❌     |     ✅      |           ✅           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Ovis-Image**           |     ❌     |     ✅      |           ❌           |       ✅        |         ❌         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -162,7 +162,6 @@ The following tables show which models support each feature:
 | **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2.5**                  |     ❌     | ❓ (one-stage) | ❓ (Ulysses only) | ✅ (Full only) |         ❓         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |    ❓ (FP8)     |        ❌         |
-| **SANA-Video-2B T2V I2V**   |     ❓     |     ❓      |           ❓           |       ❓        |         ❓         |         ❓         |   ❓    |             ❓             |          ❓           |       ❓        |        ❌         |
 | **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ✅*        |
 | **HunyuanVideo-1.5 T2V I2V** |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
 | **DreamID-Omni**             |     ❌     |     ❌      |           ❌           |       ✅        |         ❌         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -183,8 +182,8 @@ The following tables show which models support each feature:

 **Frame Interpolation Support**

- **Supported**: Wan2.2 text-to-video, image-to-video, and TI2V pipelines; SANA-Video-2B native text-to-video and image-to-video pipelines
- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, LTX-2.5, Helios, HunyuanVideo-1.5, DreamID-Omni, SANA-WM; SANA-Video-2B Diffusers-adapter text-to-video and image-to-video pipelines
+- **Supported**: Wan2.2 text-to-video, image-to-video, and TI2V pipelines
+- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, LTX-2.5, Helios, HunyuanVideo-1.5, DreamID-Omni, SANA-WM

 ### AudioGen

@@ -4,8 +4,8 @@ Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inf


 This example demonstrates how to generate videos from images using Wan2.2,
-LTX-2, HunyuanVideo-1.5, and SANA-Video Image-to-Video models with
-vLLM-Omni's offline inference API.
+LTX-2, and HunyuanVideo-1.5 Image-to-Video models with vLLM-Omni's
+offline inference API.

 ## Local CLI Usage

@@ -66,49 +66,9 @@ python image_to_video.py \
 See the [LTX-2 recipe](../../../../recipes/LTX/LTX-2.md) for all checkpoints,
 pipeline selection, T2V, defaults, and advanced options.

-### SANA-Video-2B
-
-SANA checkpoints identify their upstream T2V pipeline in `model_index.json`.
-Select the native I2V implementation explicitly:
-
-```bash
-python image_to_video.py \
-  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  --model-class-name SanaImageToVideoPipeline \
-  --image cherry_blossom.jpg \
-  --prompt "Cherry blossoms sway in the breeze as petals drift past the camera." \
-  --negative-prompt "blurry, low quality, temporal artifacts" \
-  --height 480 \
-  --width 832 \
-  --num-frames 81 \
-  --num-inference-steps 50 \
-  --guidance-scale 6.0 \
-  --extra-body '{"motion_score": 30}' \
-  --fps 16 \
-  --seed 42 \
-  --output sana_video_i2v_480p.mp4
-```
-
-For the 720p checkpoint, use
-`Efficient-Large-Model/SANA-Video_2B_720p_diffusers` with
-`--height 704 --width 1280`. The native I2V path supports both checkpoint
-variants: the 480p checkpoint uses the Wan VAE, while the 720p checkpoint uses
-the LTX-2 Video VAE.
-
-For SANA-Video, 81 frames at 16 FPS is the standard checkpoint request
-(approximately five seconds), not minute-scale long-video generation.
-Minute-scale SANA generation requires the separate LongSANA/LongLive
-autoregressive workflow, which this pipeline does not implement.
-
-See the [SANA-Video recipe](../../../../recipes/NVIDIA/SANA-Video-2B.md) for
-online serving, backend boundaries, and the tested hardware profile.
-
 Key arguments:

- `--model`: Model ID or local path (for example Wan I2V/TI2V, LTX-2, or
-  SANA-Video).
- `--model-class-name`: Explicit pipeline override. SANA-Video I2V requires
-  `SanaImageToVideoPipeline`.
+- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V, or LTX-2).
 - `--image`: Path to input image (required).
 - `--prompt`: Text description of desired motion/animation.
 - `--height/--width`: Output resolution (auto-calculated from image if not set).
@@ -16,8 +16,6 @@ For backend selection and SageAttention usage, see the [Diffusion Attention Back
 | `Lightricks/LTX-2` | 512x768 | 121 | 40 | video 3.0 / audio 7.0 | Model-dependent |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | 480x832 | 121 | 50 | 6.0 | 1×A100 80GB |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
-| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | 480x832 | 81 | 50 | 6.0 | BF16 DiT + FP32 Wan VAE |
-| `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | 704x1280 | 81 | 50 | 6.0 | ~23.6 GiB |

 ## Local CLI Usage

@@ -124,31 +122,6 @@ python text_to_video.py \
  --output quick_test.mp4
 ```

-### SANA-Video-2B
-
-```bash
-python text_to_video.py \
-  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  --model-class-name SanaVideoPipeline \
-  --prompt "A cinematic tracking shot of a sailboat crossing the ocean at sunset." \
-  --height 480 \
-  --width 832 \
-  --num-frames 81 \
-  --num-inference-steps 50 \
-  --guidance-scale 6.0 \
-  --extra-body '{"motion_score": 30}' \
-  --fps 16 \
-  --output sana_video_480p.mp4
-```
-
-For the 720p checkpoint, switch the model to
-`Efficient-Large-Model/SANA-Video_2B_720p_diffusers`, use
-`--height 704 --width 1280`, and write to a different output path. SANA-Video
-accepts `motion_score`, `clean_caption`, and `use_resolution_binning` through
-`--extra-body`. For image-to-video, use the shared
-`examples/offline_inference/image_to_video/image_to_video.py` example with
-`--model-class-name SanaImageToVideoPipeline` and `--image <path>`.
-
 ## Key Arguments

 ### Common
@@ -17,7 +17,7 @@ Shared plumbing (corpus/WAV IO, tokenizer loading, the CFG cond/uncond
 pair contract) lives in `common.py`; each task script keeps only its own
 prompt recipe and flow. Note `text_to_audio.py` here drives the Audex
 two-stage AR pipeline — for diffusion text-to-audio models (Stable Audio
-Open) use the generic
+Open, AudioX) use the generic
 `examples/offline_inference/text_to_audio/text_to_audio.py` instead.

 ## text_to_speech.py — text → speech
@@ -7,7 +7,7 @@ thinker generates interleaved 4-codebook <audiocodec_N> RVQ tokens under an
 RVQ phase mask, and the external XCodec1 checkpoint decodes them to a 16 kHz
 mono WAV.

-(For DIFFUSION text-to-audio models — Stable Audio Open — use
+(For DIFFUSION text-to-audio models — Stable Audio Open, AudioX — use
 ``examples/offline_inference/text_to_audio/text_to_audio.py`` instead; Audex
 TTA is a two-stage autoregressive pipeline with a different interface.)

@@ -1,7 +1,7 @@
 # Image-To-Video

 This shared example generates videos from images with VACE, Wan2.2, LTX-2,
-HunyuanVideo-1.5, SANA-Video, Cosmos3, and other compatible pipelines.
+HunyuanVideo-1.5, Cosmos3, and other compatible pipelines.

 - `image_to_video.py`: command-line script for single video generation with advanced options.

@@ -27,8 +27,6 @@ This folder provides a unified CLI script for image-to-video generation using vL
 | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 480 x 832 | 81 | 50 | 4.0 | Around 20–25 GiB BF16, smallest I2V model |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v` | 480 x 832 | 121 | 50 | 6.0 | Around 100 GiB at default settings; the example enables `--enable-cpu-offload` + VAE tiling/slicing to fit an 80 GiB card |
 | `Lightricks/LTX-2` | 512 x 768 | 121 | 40 | video 3.0 / audio 7.0 | Memory use depends on frame count and tensor parallelism |
-| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | 480 x 832 | 81 | 50 | 6.0 | Native `SanaImageToVideoPipeline`; Wan VAE |
-| `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | 704 x 1280 | 81 | 50 | 6.0 | Native `SanaImageToVideoPipeline`; LTX-2 Video VAE |

 !!! info
    Peak VRAM: based on basic single-card usage, batch size = 1, without any acceleration/optimization features. Some model weights cannot fit into one card with 80 GiB VRAM, which may need to use CPU offloading.
@@ -229,48 +227,6 @@ python image_to_video.py \
 See the [LTX-2 recipe](../../../recipes/LTX/LTX-2.md) for all checkpoints,
 pipeline selection, T2V, defaults, and advanced options.

-### SANA-Video-2B
-
-The checkpoint metadata selects the upstream T2V pipeline, so pass
-`SanaImageToVideoPipeline` explicitly for I2V:
-
-```bash
-python image_to_video.py \
-  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  --model-class-name SanaImageToVideoPipeline \
-  --image cherry_blossom.jpg \
-  --prompt "Cherry blossoms sway in the breeze as petals drift past the camera." \
-  --negative-prompt "blurry, low quality, temporal artifacts" \
-  --height 480 \
-  --width 832 \
-  --num-frames 81 \
-  --num-inference-steps 50 \
-  --guidance-scale 6.0 \
-  --extra-body '{"motion_score": 30}' \
-  --fps 16 \
-  --seed 42 \
-  --output sana_video_i2v_480p.mp4
-```
-
-Use the 720p checkpoint with `--height 704 --width 1280`. In this example,
-81 frames at 16 FPS means the standard approximately five-second SANA-Video
-checkpoint request. It does not enable the separate LongSANA/LongLive
-minute-scale autoregressive workflow.
-
-The native `SanaImageToVideoPipeline` supports both checkpoint variants. The
-480p path uses vLLM-Omni's `DistributedAutoencoderKLWan` wrapper and the 720p
-path uses `DistributedAutoencoderKLLTX2Video`; both wrap their corresponding
-Diffusers VAE implementations. The native denoising loop intentionally keeps
-Diffusers' checkpoint-compatible `DPMSolverMultistepScheduler`.
-
-The Diffusers adapter has completed end-to-end I2V serving validation at both
-480p and 720p. For online adapter serving, select the matching checkpoint and
-use `WIDTH=1280 HEIGHT=704` for the 720p request. Adapter T2V is also validated
-at both resolutions; see the recipe for the exact backend matrix.
-
-See the [SANA-Video recipe](../../../recipes/NVIDIA/SANA-Video-2B.md) for
-online serving, native/adapter validation boundaries, and hardware notes.
-
 ## Advanced Features

 ### CFG Parallel
@@ -3,14 +3,13 @@

 """
 Image-to-Video generation example using Wan2.2 I2V/TI2V models, LTX2/LTX-2.3,
-HunyuanVideo-1.5, SANA-Video, Cosmos3, or Wan2.1 VACE.
+HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE.

 Supports:
 - Wan2.2-I2V-A14B-Diffusers: MoE model with CLIP image encoder
 - Wan2.2-TI2V-5B-Diffusers: Unified T2V+I2V model (dense 5B)
 - LTX2 image-to-video pipeline
 - HunyuanVideo-1.5 I2V: SigLIP + VAE dual image conditioning
- SANA-Video 2B: first-frame latent conditioning at 480p or 720p
 - Wan2.1 VACE: first/last-frame, inpainting, and reference conditioning

 Usage:
@@ -39,13 +38,6 @@ Usage:
        --image input.jpg --prompt "A cat playing with yarn" \
        --flow-shift 5.0 --guidance-scale 6.0

-    # SANA-Video 2B I2V (480p)
-    python image_to_video.py --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-        --model-class-name SanaImageToVideoPipeline \
-        --image input.jpg --prompt "A cat turns toward the camera." \
-        --height 480 --width 832 --num-frames 81 --num-inference-steps 50 \
-        --guidance-scale 6.0
-
    # Cosmos3 I2V (image conditioning)
    python image_to_video.py --model nvidia/Cosmos3-Nano \
        --image input.jpg --prompt "The scene comes to life with smooth, natural motion." \
@@ -115,16 +107,13 @@ def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description=(
            "Generate a video from one or more images "
-            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, SANA-Video, Cosmos3, or Wan2.1 VACE)."
+            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE)."
        )
    )
    parser.add_argument(
        "--model",
        default="Wan-AI/Wan2.2-I2V-A14B-Diffusers",
-        help=(
-            "Diffusers I2V model ID or local path "
-            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, SANA-Video, Cosmos3, or Wan2.1 VACE)."
-        ),
+        help="Diffusers I2V model ID or local path (Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE).",
    )
    parser.add_argument(
        "--model-class-name",
@@ -388,7 +377,6 @@ def main():
    is_ltx2_distilled = "distilled" in model_class_name_lower or "distilled" in model_name
    is_ltx23 = "ltx23" in model_class_name_lower or "ltx-2.3" in model_name
    is_ltx2 = is_ltx2_distilled or is_ltx23 or "ltx2" in model_class_name_lower or "ltx-2" in model_name
-    is_sana = model_class_name == "SanaImageToVideoPipeline" or "sana-video" in model_name
    is_cosmos = "cosmos" in model_name or (model_class_name is not None and "cosmos" in model_class_name.lower())
    is_cosmos_edge = is_cosmos and ("edge" in model_name or "edge" in model_class_name_lower)

@@ -446,17 +434,6 @@ def main():
            512 * 768,
            32,
        )
-    elif is_sana:
-        is_720p = "720p" in model_name
-        d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = (
-            16,
-            6.0,
-            81,
-            50,
-            5.0,
-            (704 * 1280) if is_720p else (480 * 832),
-            32 if is_720p else 16,
-        )
    else:  # Wan2.2 / HunyuanVideo-1.5
        d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = 16, 5.0, 81, 50, 5.0, 480 * 832, 16

@@ -0,0 +1,117 @@
+"""
+Offline inference example: MammothModa2 image summarization (single AR stage).
+
+Example:
+  uv run python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \
+    --model path/to/MammothModa2-Preview \
+    --deploy-config vllm_omni/deploy/mammoth_moda2_ar.yaml \
+    --image /path/to/input.jpg \
+    --question "Please summarize the content of this image."
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+
+from PIL import Image
+from vllm import SamplingParams
+from vllm.multimodal.image import convert_image_mode
+
+from vllm_omni import Omni
+
+DEFAULT_SYSTEM = "You are a helpful assistant."
+DEFAULT_QUESTION = "Please summarize the content of this image."
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="MammothModa2 image summarization (offline, AR only).")
+    parser.add_argument("--model", type=str, required=True, help="Path to model directory or model id.")
+    parser.add_argument("--deploy-config", type=str, required=True, help="Path to deploy config yaml.")
+    parser.add_argument("--image", type=str, required=True, help="Path to input image.")
+    parser.add_argument("--question", type=str, default=DEFAULT_QUESTION, help="Question/instruction for the model.")
+    parser.add_argument("--system", type=str, default=DEFAULT_SYSTEM, help="System prompt.")
+    parser.add_argument(
+        "--max-tokens",
+        type=int,
+        default=512,
+        help="Max new tokens to generate.",
+    )
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top-p", type=float, default=0.9)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    parser.add_argument(
+        "--enable-diffusion-pipeline-profiler",
+        action="store_true",
+        help="Enable diffusion pipeline profiler to display stage durations.",
+    )
+    return parser.parse_args()
+
+
+def build_prompt(system: str, question: str) -> str:
+    return (
+        f"<|im_start|>system\n{system}<|im_end|>\n"
+        "<|im_start|>user\n"
+        "<|vision_start|><|image_pad|><|vision_end|>"
+        f"{question}<|im_end|>\n"
+        "<|im_start|>assistant\n"
+    )
+
+
+def main() -> None:
+    args = parse_args()
+
+    if not os.path.exists(args.image):
+        raise FileNotFoundError(f"Image file not found: {args.image}")
+
+    pil_image = Image.open(args.image)
+    image_data = convert_image_mode(pil_image, "RGB")
+    prompt = build_prompt(args.system, args.question)
+
+    omni = Omni(
+        model=args.model,
+        deploy_config=args.deploy_config,
+        trust_remote_code=args.trust_remote_code,
+        enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,
+    )
+    try:
+        sp = SamplingParams(
+            temperature=float(args.temperature),
+            top_p=float(args.top_p),
+            top_k=-1,
+            max_tokens=int(args.max_tokens),
+            seed=int(args.seed),
+            detokenize=True,
+        )
+        # NOTE: omni.generate() returns a Generator[OmniRequestOutput, None, None].
+        # Consume it inside the try block so the worker isn't closed early.
+        outputs = list(
+            omni.generate(
+                [
+                    {
+                        "prompt": prompt,
+                        "multi_modal_data": {"image": image_data},
+                        "additional_information": {"omni_task": ["chat"]},
+                    }
+                ],
+                [sp],
+            )
+        )
+    finally:
+        omni.close()
+
+    lines: list[str] = []
+    for stage_outputs in outputs:
+        ro = stage_outputs
+        text = ro.outputs[0].text if getattr(ro, "outputs", None) else str(ro)
+        lines.append(f"request_id: {getattr(ro, 'request_id', 'unknown')}\n")
+        lines.append("answer:\n")
+        lines.append(text.strip() + "\n")
+        lines.append("\n")
+
+    print("\n".join(lines))
+
+
+if __name__ == "__main__":
+    main()
@@ -1,10 +1,11 @@
 # Text-To-Audio

-A unified script for text-to-audio generation. Supported models:
+A unified script for text/video-to-audio generation. Supported models:

 | Model | Tasks | Notes |
 |-------|-------|-------|
 | `stabilityai/stable-audio-open-1.0` | text-to-audio | gated; uses `--audio-length` |
+| `zhangj1an/AudioX` | `t2a` / `t2m` / `v2a` / `v2m` / `tv2a` / `tv2m` | pass `--task`; video tasks need `--video` |

 The `stabilityai/stable-audio-open-1.0` pipeline generates audio from text prompts.

@@ -49,12 +50,42 @@ python text_to_audio.py \
  --output stable_audio_output.wav
 ```

+### AudioX
+
+AudioX supports six tasks. Sampler and reference knobs (declared in
+`vllm_omni/model_extras/audiox.py`) are passed via the generic `--extra-body`
+JSON flag, routed into sampling `extra_args`.
+
+Text tasks (`t2a` / `t2m`):
+
+```bash
+python text_to_audio.py \
+  --model zhangj1an/AudioX --task t2a \
+  --prompt "Fireworks burst twice, followed by a clock ticking." \
+  --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0 --seed 42 \
+  --extra-body '{"sigma_min": 0.03, "sigma_max": 1000.0}' \
+  --output t2a.wav
+```
+
+Video-conditioned tasks (`v2a` / `v2m` / `tv2a` / `tv2m`) require `--video`:
+
+```bash
+python text_to_audio.py \
+  --model zhangj1an/AudioX --task tv2a \
+  --prompt "drum beating sound and human talking" \
+  --video https://zeyuet.github.io/AudioX/static/samples/V2M/1XeBotOFqHA.mp4 \
+  --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0 \
+  --output tv2a.wav
+```
+
 Key arguments:

 - `--prompt`: text description (string).
- `--audio-start`: audio start offset in seconds (→ `audio_start_in_s`).
- `--audio-length`: audio duration in seconds (audio length for Stable Audio).
- `--extra-body`: JSON dict of model-specific knobs, merged into sampling `extra_args`.
+- `--task`: [AudioX] one of `t2a`/`t2m`/`v2a`/`v2m`/`tv2a`/`tv2m`.
+- `--video`: [AudioX `v2*`/`tv2*`] video file/URL for conditioning (→ `video_path`).
+- `--audio-start`: audio start offset in seconds (→ `audio_start_in_s` for Stable Audio, `seconds_start` for AudioX).
+- `--audio-length`: audio duration in seconds (audio length for Stable Audio, `seconds_total` for AudioX).
+- `--extra-body`: JSON dict of model-specific knobs (declared in `vllm_omni/model_extras/`), merged into sampling `extra_args`. For AudioX, sampler/reference knobs go here, e.g. `'{"sigma_min": 0.03, "sigma_max": 1000.0, "cfg_rescale": 0.0, "audio_path": "ref.wav"}'`.
 - `--negative-prompt`: negative prompt for classifier-free guidance.
 - `--seed`: integer seed for deterministic generation.
 - `--guidance-scale`: classifier-free guidance scale.
@@ -2,13 +2,15 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 """
-Example script for text-to-audio generation with diffusion audio models.
+Example script for text/video-to-audio generation with diffusion audio models.

 This script supports:
  * Stable Audio Open (`stabilityai/stable-audio-open-1.0`) — text-to-audio.
+  * AudioX (`zhangj1an/AudioX`) — six tasks: t2a / t2m / v2a / v2m / tv2a / tv2m.

-Model-specific generation knobs can be passed via `--extra-body` and are
-merged into `sampling_params.extra_args`.
+Model-specific generation knobs are declared in `vllm_omni/model_extras` and
+routed into `sampling_params.extra_args` via `apply_declared_extra_args`, so
+the same script serves both models without per-model branching in the engine.

 Usage:
    # Stable Audio Open
@@ -16,6 +18,15 @@ Usage:
    python text_to_audio.py --prompt "A piano playing a gentle melody" --audio-length 10.0
    python text_to_audio.py --prompt "Thunder and rain sounds" --negative-prompt "Low quality"
    python text_to_audio.py --prompt "A soft synth pad" --cache-backend tea_cache
+
+    # AudioX (text-to-audio / text-to-music)
+    python text_to_audio.py --model zhangj1an/AudioX --task t2a \
+        --prompt "Fireworks burst twice, then a clock ticking." \
+        --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0
+    # AudioX video-conditioned (v2*/tv2*) — pass a video file/URL
+    python text_to_audio.py --model zhangj1an/AudioX --task tv2a \
+        --prompt "drum beating sound and human talking" --video clip.mp4 \
+        --extra-body '{"sigma_min": 0.03, "sigma_max": 1000.0}'
 """

 import argparse
@@ -27,11 +38,16 @@ from typing import Any
 import numpy as np
 import torch

+from vllm_omni.diffusion.utils.param_utils import apply_declared_extra_args
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.model_extras import get_extra_body_params, get_model_class_name
 from vllm_omni.outputs import OmniRequestOutput
 from vllm_omni.platforms import current_omni_platform

+AUDIOX_TASKS = ("t2a", "t2m", "v2a", "v2m", "tv2a", "tv2m")
+AUDIOX_VIDEO_TASKS = frozenset({"v2a", "v2m", "tv2a", "tv2m"})
+

 def parse_extra_body(value: str) -> dict[str, Any]:
    """Parse a JSON object string for --extra-body."""
@@ -49,7 +65,7 @@ def parse_args() -> argparse.Namespace:
    parser.add_argument(
        "--model",
        default="stabilityai/stable-audio-open-1.0",
-        help="Audio diffusion model name or local path. Supported: stabilityai/stable-audio-open-1.0.",
+        help="Audio diffusion model name or local path. Supported: stabilityai/stable-audio-open-1.0, zhangj1an/AudioX.",
    )
    parser.add_argument(
        "--prompt",
@@ -80,13 +96,19 @@ def parse_args() -> argparse.Namespace:
        "--audio-start",
        type=float,
        default=0.0,
-        help="Audio start offset in seconds (maps to audio_start_in_s).",
+        help=(
+            "Audio start offset in seconds. Applies to both models "
+            "(audio_start_in_s for Stable Audio, seconds_start for AudioX)."
+        ),
    )
    parser.add_argument(
        "--audio-length",
        type=float,
        default=10.0,
-        help=("Audio duration in seconds. Maps to audio length for Stable Audio (max ~47s for stable-audio-open-1.0)."),
+        help=(
+            "Audio duration in seconds. Maps to audio length for Stable Audio "
+            "(max ~47s for stable-audio-open-1.0) and seconds_total for AudioX."
+        ),
    )
    parser.add_argument(
        "--num-inference-steps",
@@ -100,20 +122,41 @@ def parse_args() -> argparse.Namespace:
        default=1,
        help="Number of audio waveforms to generate for the given prompt.",
    )
+    parser.add_argument(
+        "--task",
+        type=str,
+        default=None,
+        choices=list(AUDIOX_TASKS),
+        help=(
+            "[AudioX only] Generation task: t2a/t2m/v2a/v2m/tv2a/tv2m. "
+            "v2*/tv2* require --video. Ignored for Stable Audio Open."
+        ),
+    )
+    parser.add_argument(
+        "--video",
+        type=str,
+        default=None,
+        help="[AudioX v2*/tv2*] Video file path or URL used for conditioning.",
+    )
    parser.add_argument(
        "--extra-body",
        type=parse_extra_body,
        default=None,
        help=(
-            "Model-specific generation params as a JSON object, merged into "
-            "sampling extra_args. Values here override the equivalent flags."
+            "Model-specific generation params as a JSON object, e.g. "
+            '\'{"sigma_min": 0.03, "sigma_max": 1000.0}\'. Each key is filtered '
+            "against the model's declared extra_body_params (see vllm_omni/model_extras), "
+            "so unknown keys are silently dropped. Values here override the equivalent flags."
        ),
    )
    parser.add_argument(
        "--output",
        type=str,
        default=None,
-        help=("Path to save the generated audio (WAV format). Defaults to stable_audio_output.wav."),
+        help=(
+            "Path to save the generated audio (WAV format). Defaults to "
+            "audiox_output.wav for AudioX and stable_audio_output.wav otherwise."
+        ),
    )
    parser.add_argument(
        "--sample-rate",
@@ -232,6 +275,8 @@ def save_audio(audio_data: np.ndarray, output_path: str, sample_rate: int = 4410

 def main():
    args = parse_args()
+    model_name = str(args.model).lower() if args.model else ""
+    is_audiox = "audiox" in model_name
    generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(args.seed)
    cache_config = None
    if args.cache_backend == "tea_cache":
@@ -240,9 +285,11 @@ def main():
        }

    print(f"\n{'=' * 60}")
-    print("Text-to-Audio Generation")
+    print("Text/Video-to-Audio Generation")
    print(f"{'=' * 60}")
    print(f"  Model: {args.model}")
+    if args.task is not None:
+        print(f"  Task: {args.task}")
    print(f"  Prompt: {args.prompt}")
    print(f"  Negative prompt: {args.negative_prompt}")
    print(f"  Duration: {args.audio_length}s")
@@ -258,6 +305,8 @@ def main():
    print(f"  Seed: {args.seed}")
    print(f"{'=' * 60}\n")

+    # Initialize Omni. AudioX requires an explicit pipeline class; Stable Audio
+    # is resolved from its config.
    omni_kwargs: dict[str, Any] = dict(
        model=args.model,
        use_hsdp=args.use_hsdp,
@@ -269,7 +318,11 @@ def main():
        enable_cpu_offload=args.enable_cpu_offload,
        enable_layerwise_offload=args.enable_layerwise_offload,
    )
+    if is_audiox:
+        omni_kwargs["model_class_name"] = "AudioXPipeline"
    omni = Omni(**omni_kwargs)
+    model_class_name = get_model_class_name(omni)
+    declared_extra_body_params = get_extra_body_params(model_class_name)

    diffusion_params = OmniDiffusionSamplingParams(
        generator=generator,
@@ -279,19 +332,38 @@ def main():
        seed=args.seed,
    )

-    # Negative prompt is opt-in (off by default). Including a
+    # Negative prompt is opt-in for both models (off by default). Including a
    # non-empty negative prompt enables the separate CFG conditioning branch.
    prompt: dict[str, Any] = {"prompt": args.prompt}
    if args.negative_prompt:
        prompt["negative_prompt"] = args.negative_prompt

-    audio_end_in_s = args.audio_start + args.audio_length
-    diffusion_params.extra_args = {
-        "audio_start_in_s": args.audio_start,
-        "audio_end_in_s": audio_end_in_s,
-    }
-    if args.extra_body:
-        diffusion_params.extra_args.update(args.extra_body)
+    if model_class_name == "AudioXPipeline":
+        if args.task is None:
+            raise ValueError("AudioX requires --task (one of t2a/t2m/v2a/v2m/tv2a/tv2m).")
+        if args.task in AUDIOX_VIDEO_TASKS and not args.video:
+            raise ValueError(f"AudioX task {args.task!r} requires --video.")
+        # AudioX core knobs are surfaced as shared flags; sampler/reference knobs
+        # (sigma_min/sigma_max/cfg_rescale/audio_path) arrive via --extra-body and
+        # are filtered against the declared extra_body contract.
+        user_extra: dict[str, Any] = {
+            "audiox_task": args.task,
+            "seconds_start": args.audio_start,
+            "seconds_total": args.audio_length,
+            "video_path": args.video,
+        }
+        if args.extra_body:
+            user_extra.update(args.extra_body)
+        apply_declared_extra_args(diffusion_params, declared_extra_body_params, user_extra)
+    else:
+        # Stable Audio Open path (backward-compatible).
+        audio_end_in_s = args.audio_start + args.audio_length
+        diffusion_params.extra_args = {
+            "audio_start_in_s": args.audio_start,
+            "audio_end_in_s": audio_end_in_s,
+        }
+        if args.extra_body:
+            diffusion_params.extra_args.update(args.extra_body)

    generation_start = time.perf_counter()

@@ -302,7 +374,7 @@ def main():

    print(f"Total generation time: {generation_time:.2f} seconds")

-    output = args.output or "stable_audio_output.wav"
+    output = args.output or ("audiox_output.wav" if model_class_name == "AudioXPipeline" else "stable_audio_output.wav")
    output_path = Path(output)
    output_path.parent.mkdir(parents=True, exist_ok=True)
    suffix = output_path.suffix or ".wav"
@@ -89,7 +89,7 @@ def parse_args() -> argparse.Namespace:
        "Qwen/Qwen-Image, Tongyi-MAI/Z-Image-Turbo, Qwen/Qwen-Image-2512, stepfun-ai/NextStep-1.1, "
        "black-forest-labs/FLUX.1-dev, black-forest-labs/FLUX.2-klein-9B, "
        "black-forest-labs/FLUX.2-dev, tencent/HunyuanImage-3.0-Instruct, "
-        "HiDream-ai/HiDream-O1-Image, meituan-longcat/LongCat-Image, OvisAI/Ovis-Image, "
+        "meituan-longcat/LongCat-Image, OvisAI/Ovis-Image, "
        "stabilityai/stable-diffusion-3.5-medium, Tongyi-MAI/Z-Image-Turbo and etc.",
    )
    parser.add_argument(
@@ -599,6 +599,17 @@ def main():
            if args.seed is not None and hasattr(params, "seed"):
                params.seed = args.seed

+            # MammothModa2's AR stage emits one visual token per grid cell,
+            # one EOL token per row, and one final look-ahead token whose hidden
+            # state is unavailable. Size the first stage from the prompt metadata
+            # instead of SamplingParams' default of 16 tokens.
+            prompt_info = prompt_dict.get("additional_information", {})
+            if idx == 0 and prompt_info.get("omni_task") == ["t2i"]:
+                ar_width = int(prompt_info.get("ar_width", [0])[0])
+                ar_height = int(prompt_info.get("ar_height", [0])[0])
+                if ar_width > 0 and ar_height > 0:
+                    params.max_tokens = ar_height * (ar_width + 1) + 1
+
    if not diffusion_replaced and len(sampling_params_list) == 1:
        sampling_params_list = [diffusion_params]

@@ -420,7 +420,8 @@ against transformers 5.x, so nothing is lost by holding it there.

 (The same clash is why this is documented here rather than declared in
 `pyproject.toml`: an extra resolves together with the base dependencies, so
-`vllm-omni[gepard]` would be unsatisfiable rather than opt-in.)
+`vllm-omni[gepard]` would be unsatisfiable rather than opt-in. SoulX-Singer English
+SVS documents `nemo_toolkit[asr]` the same way.)

 On a host whose CUDA toolkit cannot build kernels for the GPU — no `nvcc`/`ninja`, or a
 consumer Blackwell (`sm_120`) card — also set `VLLM_USE_FLASHINFER_SAMPLER=0`. vLLM's
@@ -616,3 +617,100 @@ Available voice presets are listed on the HF model card (`mistralai/Voxtral-4B-T
 - `--num-prompts N` replicates the prompt for performance measurement.
 - `--concurrency M` requires `--streaming` and must evenly divide `--num-prompts`.
 - Run `--help` for the full argument surface.
+
+---
+
+## SoulX-Singer
+
+Singing voice synthesis (SVS) and conversion (SVC) at 24 kHz. Script: `soulxsinger/end2end.py`. Deploy: `vllm_omni/deploy/soulxsinger_svs.yaml` or `soulxsinger_svc.yaml`.
+
+### Prerequisites
+
+Download DiT and preprocess weights, then set up separate SVS / SVC view directories. Copy `soulxsinger/utils/phoneme/phone_set.json` from upstream [SoulX-Singer](https://github.com/Soul-AILab/SoulX-Singer) into the model weights dir as `phoneme/phone_set.json` — HuggingFace does not ship it.
+
+```bash
+# 1. DiT weights
+export BASE=path/to/SoulX-Singer
+export PREPROCESS=path/to/SoulX-Singer-Preprocess
+export SVC_DIR=path/to/SoulX-Singer-svc
+
+huggingface-cli download Soul-AILab/SoulX-Singer --local-dir "$BASE"
+
+# 2. Preprocess weights (required)
+huggingface-cli download Soul-AILab/SoulX-Singer-Preprocess --local-dir "$PREPROCESS"
+export SOULX_PREPROCESS_WEIGHTS_DIR="$PREPROCESS"
+
+# 3. SVS / SVC view directories
+mkdir -p "$SVC_DIR"
+cp $BASE/{config.yaml,README.md,assets} $SVC_DIR
+mv $BASE/model-svc.pt $SVC_DIR/model-svc.pt
+
+cat > "$BASE/config.json" <<'EOF'
+{
+  "model_type": "soulxsinger",
+  "architectures": ["SoulXSingerPipeline"],
+  "max_num_seqs": 1
+}
+EOF
+
+cat > "$SVC_DIR/config.json" <<'EOF'
+{
+  "model_type": "soulxsinger",
+  "architectures": ["SoulXSingerSVCPipeline"],
+  "max_num_seqs": 1
+}
+EOF
+```
+
+`config.yaml` hyper-parameters live under `$BASE`; each view's `config.json` `architectures` field is the single source of truth for SVS vs SVC. Point `--model` at the matching directory (`$BASE` for SVS, `$SVC_DIR` for SVC). Deploy YAML is chosen automatically from `config.json`; optional `--svs` / `--svc` only assert the mode matches.
+
+**Online preprocess** is the default: pass `--prompt-audio` and `--target-audio`, and the worker runs vocal separation, F0, and (for SVS) lyrics/MIDI before DiT. Install only what your run needs:
+
+```bash
+pip install "BS-RoFormer"   # vocal sep + F0 on GPU — SVS and SVC
+```
+
+Mandarin SVS also needs FunASR and Chinese G2P; `ffmpeg` must be on `PATH`:
+
+```bash
+# install optional dependencies:
+pip install -e ".[soulx-svs]"
+```
+
+English SVS adds NeMo ASR and NLTK data; pass `--language English`:
+
+```bash
+pip install "nemo_toolkit[asr]==2.6.1" lhotse==1.32.2
+python -c "import nltk; nltk.download('cmudict'); nltk.download('averaged_perceptron_tagger_eng')"
+```
+
+**Precomputed metadata** is the alternative: pass both `--prompt-metadata-path` and `--target-metadata-path` and skip online ASR/ROSVOT — none of the packages above are required. JSON can be produced by integrated preprocess on a prior run, or by upstream [SoulX-Singer](https://github.com/Soul-AILab/SoulX-Singer) `preprocess/` scripts if you prefer to run that outside vLLM-Omni.
+
+### Quick start
+
+```bash
+# SVS — default demo audio: tests/assets/soulxsinger/zh_prompt.mp3 + music.mp3
+python examples/offline_inference/text_to_speech/soulxsinger/end2end.py \
+    --model "$BASE" \
+    --preprocess-weights-dir "$PREPROCESS" \
+    --control score \
+    --num-inference-steps 32 \
+    -o output.wav
+
+python examples/offline_inference/text_to_speech/soulxsinger/end2end.py \
+    --model "$SVC_DIR" \
+    --preprocess-weights-dir "$PREPROCESS" \
+    --svc \
+    --num-inference-steps 32 \
+    -o output_svc.wav
+```
+
+`SOULX_PREPROCESS_WEIGHTS_DIR` makes `--preprocess-weights-dir` optional. Long SVS targets are handled in one request. See `end2end.py --help` for `--pitch-shift`, `--vocal-sep`, `--auto-shift`, and language/control options.
+
+### Notes
+
+- Output: 24 kHz mono WAV; batch only.
+- Defaults match upstream: `--guidance-scale 3.0`, `--seed 42`, `--auto-shift` on.
+- SVS `--control`: `score` or `melody`. MIDI / lyric QC: upstream `midi_editor` only.
+
+---
@@ -0,0 +1,162 @@
+"""Repeatable SoulX-Singer offline benchmark (preprocess + DiT, RTF + stage timings).
+
+Usage:
+    python benchmark.py --model /path/to/SoulX-Singer --svs \\
+        --prompt-audio ../SoulX-Singer/example/audio/zh_prompt.mp3 \\
+        --target-audio ../SoulX-Singer/example/audio/music.mp3 \\
+        --preprocess-weights-dir /path/to/SoulX-Singer-Preprocess \\
+        -o benchmark.wav
+
+    python benchmark.py ... --enable-diffusion-pipeline-profiler
+"""
+
+import argparse
+import statistics
+import time
+from pathlib import Path
+
+import soundfile as sf
+from end2end import (
+    SVC_DEPLOY_CONFIG,
+    SVS_DEPLOY_CONFIG,
+    add_inference_args,
+    build_sampling,
+    extract_audio,
+    resolve_preprocess_weights_dir,
+)
+
+from vllm_omni.entrypoints.omni import Omni
+
+
+def _audio_duration_sec(outputs) -> float:
+    audio_np, sr = extract_audio(outputs)
+    return audio_np.size / sr
+
+
+def _stage_durations_from_output(outputs) -> dict[str, float]:
+    omni_out = outputs[0]
+    durations = getattr(omni_out, "stage_durations", None) or {}
+    if durations:
+        return dict(durations)
+    ro = omni_out
+    if ro is not None:
+        return dict(getattr(ro, "stage_durations", None) or {})
+    return {}
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="SoulX-Singer offline benchmark")
+    add_inference_args(parser)
+    parser.add_argument("--warmup", type=int, default=1, help="Warmup runs (excluded from stats)")
+    parser.add_argument("--runs", type=int, default=3, help="Measured runs")
+    parser.add_argument(
+        "--enable-diffusion-pipeline-profiler",
+        action="store_true",
+        help="Enable low-overhead stage timing in DiffusionOutput.stage_durations",
+    )
+    parser.add_argument(
+        "--enforce-eager",
+        action="store_true",
+        help="Disable torch.compile on diff_estimator (eager baseline for A/B)",
+    )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        default=None,
+        choices=["bfloat16", "bf16", "float16", "fp16", "half"],
+        help="DiT trunk dtype (default from deploy YAML). Use float16 to match upstream --fp16.",
+    )
+    parser.add_argument(
+        "--output",
+        "-o",
+        type=str,
+        default=None,
+        help="Save WAV from the last measured run (after warmup)",
+    )
+    parser.add_argument(
+        "--save-each-run",
+        action="store_true",
+        help="With --output, also write measured runs as <stem>_run1.wav, _run2.wav, ...",
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    preprocess_dir = resolve_preprocess_weights_dir(args.preprocess_weights_dir)
+    deploy_config = args.deploy_config or str(SVS_DEPLOY_CONFIG if args.svs else SVC_DEPLOY_CONFIG)
+    mode = "SVS" if args.svs else "SVC"
+
+    omni_kwargs: dict = {
+        "model": args.model,
+        "deploy_config": deploy_config,
+        "async_chunk": False,  # SoulX-Singer currently supports only batch mode (pseudo-streaming was stashed)
+    }
+    if args.enable_diffusion_pipeline_profiler:
+        omni_kwargs["enable_diffusion_pipeline_profiler"] = True
+    if args.enforce_eager:
+        omni_kwargs["enforce_eager"] = True
+    if args.dtype is not None:
+        omni_kwargs["dtype"] = args.dtype
+
+    compile_mode = "eager" if args.enforce_eager else "torch.compile"
+    dtype_label = args.dtype or "from deploy YAML"
+    print(f"Loading SoulX-Singer {mode} from {args.model} [{compile_mode}, dtype={dtype_label}]")
+
+    omni = Omni(**omni_kwargs)
+    kind = "svs" if args.svs else "svc"
+    sampling = build_sampling(args, preprocess_weights_dir=preprocess_dir, kind=kind)
+    prompt = {"prompt_token_ids": [0]}
+
+    latencies_ms: list[float] = []
+    rtfs: list[float] = []
+    last_stages: dict[str, float] = {}
+    last_measured_outputs = None
+
+    total_runs = args.warmup + args.runs
+    for run_idx in range(total_runs):
+        is_warmup = run_idx < args.warmup
+        label = "warmup" if is_warmup else f"run {run_idx - args.warmup + 1}/{args.runs}"
+        t0 = time.perf_counter()
+        outputs = list(omni.generate([prompt], sampling))
+        elapsed_ms = (time.perf_counter() - t0) * 1000.0
+        audio_sec = _audio_duration_sec(outputs)
+        rtf = (elapsed_ms / 1000.0) / audio_sec if audio_sec > 0 else float("inf")
+        print(f"[{label}] client={elapsed_ms:.1f} ms, audio={audio_sec:.2f}s, RTF={rtf:.3f}")
+        if not is_warmup:
+            latencies_ms.append(elapsed_ms)
+            rtfs.append(rtf)
+            last_stages = _stage_durations_from_output(outputs)
+            last_measured_outputs = outputs
+            if args.output and args.save_each_run:
+                measured_idx = run_idx - args.warmup + 1
+                out_path = Path(args.output)
+                run_path = out_path.with_name(f"{out_path.stem}_run{measured_idx}{out_path.suffix}")
+                audio_np, sr = extract_audio(outputs)
+                sf.write(str(run_path), audio_np, sr)
+                print(f"  saved {run_path} ({sr} Hz, {audio_np.size / sr:.2f}s)")
+
+    omni.close()
+
+    if args.output and last_measured_outputs is not None:
+        audio_np, sr = extract_audio(last_measured_outputs)
+        out_path = Path(args.output)
+        out_path.parent.mkdir(parents=True, exist_ok=True)
+        sf.write(str(out_path), audio_np, sr)
+        print(f"\nSaved last measured run → {out_path} ({sr} Hz, {audio_np.size / sr:.2f}s)")
+
+    if latencies_ms:
+        print("\n=== Summary (measured runs) ===")
+        print(f"client_ms: mean={statistics.mean(latencies_ms):.1f}, stdev={statistics.pstdev(latencies_ms):.1f}")
+        print(f"RTF:       mean={statistics.mean(rtfs):.3f}, stdev={statistics.pstdev(rtfs):.3f}")
+        if last_stages:
+            print("\nStage durations (last measured run):")
+            for name, value in sorted(last_stages.items()):
+                if name.endswith("_ms"):
+                    print(f"  {name}: {value:.1f} ms")
+                else:
+                    print(f"  {name}: {value * 1000.0:.1f} ms")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,285 @@
+"""Offline SoulX-Singer SVS / SVC: single-stage DiT (preprocess inline).
+
+Pass raw prompt/target audio; vLLM-Omni runs integrated preprocess then inference.
+
+Usage:
+    # SVS (default)
+    python end2end.py --model /path/to/SoulX-Singer \\
+        --prompt-audio path/to/prompt/audio.mp3 \\
+        --target-audio path/to/target/audio.mp3 \\
+        --preprocess-weights-dir /path/to/SoulX-Singer-Preprocess \\
+        -o output.wav
+
+    # SVC (config.json must declare SoulXSingerSVCPipeline; --model points at SVC view)
+    python end2end.py --model /path/to/SoulX-Singer-svc \\
+        --prompt-audio path/to/prompt/audio.mp3 \\
+        --target-audio path/to/target/audio.mp3 \\
+        --preprocess-weights-dir /path/to/SoulX-Singer-Preprocess \\
+        -o output_svc.wav
+"""
+
+import argparse
+import os
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+
+from vllm_omni.diffusion.models.soulx_singer.utils import resolve_soulx_kind, validate_soulx_extra_args
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+REPO_ROOT = Path(__file__).resolve().parents[4]
+SVS_DEPLOY_CONFIG = REPO_ROOT / "vllm_omni" / "deploy" / "soulxsinger_svs.yaml"
+SVC_DEPLOY_CONFIG = REPO_ROOT / "vllm_omni" / "deploy" / "soulxsinger_svc.yaml"
+DEFAULT_PROMPT_AUDIO = REPO_ROOT / "tests" / "assets" / "soulxsinger" / "zh_prompt.mp3"
+DEFAULT_TARGET_AUDIO = REPO_ROOT / "tests" / "assets" / "soulxsinger" / "music.mp3"
+_SAMPLE_RATE = 24000
+
+
+def _require_paths(paths: dict[str, str | None]) -> None:
+    missing = [name for name, p in paths.items() if not p or not os.path.isfile(p)]
+    if missing:
+        raise FileNotFoundError("Missing file(s): " + ", ".join(f"{k}={paths[k]!r}" for k in missing))
+
+
+def resolve_preprocess_weights_dir(explicit: str | None) -> str:
+    if explicit:
+        path = Path(explicit).expanduser().resolve()
+        if not (path / "rmvpe" / "rmvpe.pt").is_file():
+            raise FileNotFoundError(f"--preprocess-weights-dir must contain rmvpe/rmvpe.pt: {path}")
+        return str(path)
+    env = os.environ.get("SOULX_PREPROCESS_WEIGHTS_DIR")
+    if env:
+        path = Path(env).expanduser().resolve()
+        if (path / "rmvpe" / "rmvpe.pt").is_file():
+            return str(path)
+    raise FileNotFoundError(
+        "SoulX preprocess weights not found. Pass --preprocess-weights-dir or set "
+        "SOULX_PREPROCESS_WEIGHTS_DIR to a directory containing rmvpe/rmvpe.pt "
+        "(huggingface-cli download Soul-AILab/SoulX-Singer-Preprocess)."
+    )
+
+
+def build_sampling(
+    args: argparse.Namespace,
+    *,
+    preprocess_weights_dir: str,
+    kind: str,
+) -> OmniDiffusionSamplingParams:
+    extra: dict = {
+        "prompt_audio": os.path.abspath(args.prompt_audio),
+        "target_audio": os.path.abspath(args.target_audio),
+        "preprocess_weights_dir": preprocess_weights_dir,
+        "vocal_sep": args.vocal_sep,
+        "auto_shift": args.auto_shift,
+        "pitch_shift": args.pitch_shift,
+    }
+    if kind == "svs":
+        extra["language"] = args.language
+        extra["control"] = args.control
+        if args.prompt_metadata_path:
+            extra["prompt_metadata_path"] = os.path.abspath(args.prompt_metadata_path)
+        if args.target_metadata_path:
+            extra["target_metadata_path"] = os.path.abspath(args.target_metadata_path)
+        if args.prompt_metadata_path or args.target_metadata_path:
+            extra["audio_path"] = os.path.abspath(args.prompt_wav_path or args.prompt_audio)
+
+    extra = validate_soulx_extra_args(kind, extra)
+
+    return OmniDiffusionSamplingParams(
+        num_inference_steps=args.num_inference_steps,
+        guidance_scale=args.guidance_scale,
+        seed=args.seed,
+        extra_args=extra,
+    )
+
+
+def extract_audio(outputs) -> tuple[np.ndarray, int]:
+    out = outputs[0]
+    mm = getattr(out, "multimodal_output", None)
+    if mm is None:
+        ro = out
+        if ro is not None:
+            mm = getattr(ro, "multimodal_output", None)
+            if mm is None and ro.outputs:
+                mm = getattr(ro.outputs[0], "multimodal_output", None)
+    if not mm or "audio" not in mm:
+        raise RuntimeError("No audio in multimodal_output")
+
+    audio_val = mm["audio"]
+    if isinstance(audio_val, list):
+        chunks = [
+            np.asarray(chunk.detach().cpu().float().numpy() if hasattr(chunk, "detach") else chunk).reshape(-1)
+            for chunk in audio_val
+            if chunk is not None
+        ]
+        audio_np = np.concatenate(chunks, axis=0) if chunks else np.array([], dtype=np.float32)
+    elif hasattr(audio_val, "cpu"):
+        audio_np = audio_val.detach().cpu().numpy().squeeze()
+    else:
+        audio_np = np.asarray(audio_val, dtype=np.float32).squeeze()
+
+    sr_val = mm.get("sr", mm.get("audio_sample_rate", _SAMPLE_RATE))
+    if isinstance(sr_val, list) and sr_val:
+        sr_val = sr_val[-1]
+    if hasattr(sr_val, "item"):
+        sr_val = sr_val.item()
+    return audio_np.astype(np.float32, copy=False), int(sr_val)
+
+
+def add_inference_args(parser: argparse.ArgumentParser) -> None:
+    mode = parser.add_mutually_exclusive_group()
+    mode.add_argument(
+        "--svc",
+        action="store_true",
+        help="Assert SVC mode (must match config.json architectures)",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="SoulX-Singer weight directory ($BASE for SVS, $SVC_DIR for SVC)",
+    )
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help="Override deploy YAML (default: soulxsinger_svs.yaml or soulxsinger_svc.yaml)",
+    )
+    parser.add_argument(
+        "--prompt-audio",
+        type=str,
+        default=str(DEFAULT_PROMPT_AUDIO),
+        help="Prompt / reference audio (mp3 or wav)",
+    )
+    parser.add_argument(
+        "--target-audio",
+        type=str,
+        default=str(DEFAULT_TARGET_AUDIO),
+        help="Target content audio (mp3 or wav)",
+    )
+    parser.add_argument(
+        "--prompt-metadata-path",
+        type=str,
+        default=None,
+        help="SVS precomputed prompt metadata.json (skips online ASR/ROSVOT for prompt)",
+    )
+    parser.add_argument(
+        "--target-metadata-path",
+        type=str,
+        default=None,
+        help="SVS precomputed target metadata.json (skips online ASR/ROSVOT for target)",
+    )
+    parser.add_argument(
+        "--prompt-wav-path",
+        type=str,
+        default=None,
+        help="SVS prompt vocal wav for metadata processor when using --prompt-metadata-path (default: --prompt-audio)",
+    )
+    parser.add_argument(
+        "--preprocess-weights-dir",
+        type=str,
+        default=None,
+        help="SoulX-Singer-Preprocess weights (or set SOULX_PREPROCESS_WEIGHTS_DIR)",
+    )
+    parser.add_argument(
+        "--language",
+        type=str,
+        default="Mandarin",
+        help="SVS lyrics language (Mandarin, Cantonese, English)",
+    )
+    parser.add_argument(
+        "--control",
+        type=str,
+        default="melody",
+        choices=["score", "melody"],
+        help="SVS control mode",
+    )
+    parser.add_argument(
+        "--vocal-sep",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+        help="Run vocal separation during preprocess",
+    )
+    parser.add_argument(
+        "--auto-shift",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Auto pitch shift to align target pitch with prompt",
+    )
+    parser.add_argument("--pitch-shift", type=int, default=0)
+    parser.add_argument("--num-inference-steps", type=int, default=32)
+    parser.add_argument("--guidance-scale", type=float, default=3.0)
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="Optional CFM RNG seed. Omit for non-deterministic sampling.",
+    )
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="SoulX-Singer offline SVS / SVC (preprocess + DiT)")
+    add_inference_args(parser)
+    parser.add_argument("--output", "-o", type=str, default="output.wav")
+    args = parser.parse_args()
+    return args
+
+
+def main() -> None:
+    args = parse_args()
+    kind = resolve_soulx_kind(args.model)
+
+    if not (args.svc ^ (kind == "svs")):
+        raise ValueError(f"mode mismatch: {args.model}/config.json (declares {kind!r}), --svc={args.svc}")
+
+    required = {
+        "prompt_audio": args.prompt_audio,
+        "target_audio": args.target_audio,
+    }
+    if kind == "svs" and bool(args.prompt_metadata_path) ^ bool(args.target_metadata_path):
+        raise ValueError(
+            "SVS precomputed metadata requires both --prompt-metadata-path and "
+            "--target-metadata-path (or neither, for online preprocess)."
+        )
+    if args.prompt_metadata_path:
+        required["prompt_metadata_path"] = args.prompt_metadata_path
+        required["target_metadata_path"] = args.target_metadata_path
+    if args.prompt_wav_path:
+        required["prompt_wav_path"] = args.prompt_wav_path
+    _require_paths(required)
+    preprocess_dir = resolve_preprocess_weights_dir(args.preprocess_weights_dir)
+    deploy_config = args.deploy_config or str(SVC_DEPLOY_CONFIG if kind == "svc" else SVS_DEPLOY_CONFIG)
+    mode = kind.upper()
+
+    print(f"Loading SoulX-Singer {mode} from {args.model}")
+    print(f"  deploy: {deploy_config}")
+    print(f"  prompt: {args.prompt_audio}")
+    print(f"  target: {args.target_audio}")
+
+    omni = Omni(model=args.model, deploy_config=deploy_config, async_chunk=False)  # SoulX-Singer is batch-only
+    sampling = build_sampling(
+        args,
+        preprocess_weights_dir=preprocess_dir,
+        kind=kind,
+    )
+    prompt = {"prompt_token_ids": [0], "multi_modal_data": {}}
+
+    print("Running generate (inline preprocess + DiT)...")
+    outputs = list(omni.generate([prompt], sampling))
+    if not outputs:
+        raise RuntimeError("No output from omni.generate")
+
+    err = getattr(outputs[0], "error", None)
+    if err is not None:
+        raise RuntimeError(f"Pipeline error: {err}")
+
+    audio_np, sr = extract_audio(outputs)
+    sf.write(args.output, audio_np, sr)
+    print(f"Saved {args.output} ({sr} Hz, {len(audio_np) / sr:.2f}s)")
+    omni.close()
+
+
+if __name__ == "__main__":
+    main()
@@ -14,8 +14,6 @@ A unified script for text-to-video generation. Supports multiple models with mod
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
 | `nvidia/Cosmos3-Nano` | 720x1280 | 189 | 35 | 6.0 | ~46 GiB (peak, 720p) |
 | `BestWishYsh/Helios-Base` / `Helios-Mid` / `Helios-Distilled` | 384x640 | 99 | 50 | 5.0 / 5.0 / 1.0 | — |
-| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | 480x832 | 81 | 50 | 6.0 | BF16 DiT + FP32 Wan VAE wrapper |
-| `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | 704x1280 | 81 | 50 | 6.0 | BF16 DiT + LTX-2 Video VAE wrapper |

 ## Local CLI Usage

@@ -153,40 +151,6 @@ python text_to_video.py \
  --output quick_test.mp4
 ```

-### SANA-Video-2B
-
-Select the native T2V pipeline explicitly:
-
-```bash
-python text_to_video.py \
-  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  --model-class-name SanaVideoPipeline \
-  --prompt "A cinematic tracking shot of a sailboat crossing the ocean at sunset." \
-  --height 480 \
-  --width 832 \
-  --num-frames 81 \
-  --num-inference-steps 50 \
-  --guidance-scale 6.0 \
-  --extra-body '{"motion_score": 30}' \
-  --fps 16 \
-  --output sana_video_480p.mp4
-```
-
-For the 720p checkpoint, switch the model to
-`Efficient-Large-Model/SANA-Video_2B_720p_diffusers`, use
-`--height 704 --width 1280`, and write to a different output path. Both native
-checkpoint variants are supported. The Diffusers adapter T2V compatibility
-path is also validated at both resolutions through online serving.
-
-The native 480p path uses vLLM-Omni's `DistributedAutoencoderKLWan` wrapper;
-the native 720p path uses `DistributedAutoencoderKLLTX2Video`. Both wrappers
-retain their underlying Diffusers autoencoders. The pipeline intentionally
-uses Diffusers' checkpoint-compatible `DPMSolverMultistepScheduler`.
-
-For image-to-video, use the shared
-[`image_to_video.py`](../image_to_video/README.md#sana-video-2b) example with
-`--model-class-name SanaImageToVideoPipeline` and `--image <path>`.
-
 ### Cosmos3

 ```bash
@@ -114,32 +114,12 @@ _MODEL_PRESETS = {
        "fps": 24,
        "output": "ltx23_output.mp4",
    },
-    "sana_480p": {
-        "height": 480,
-        "width": 832,
-        "num_frames": 81,
-        "num_inference_steps": 50,
-        "guidance_scale": 6.0,
-        "fps": 16,
-        "output": "sana_video_480p.mp4",
-    },
-    "sana_720p": {
-        "height": 704,
-        "width": 1280,
-        "num_frames": 81,
-        "num_inference_steps": 50,
-        "guidance_scale": 6.0,
-        "fps": 16,
-        "output": "sana_video_720p.mp4",
-    },
 }


 def _detect_preset(model: str, model_class_name: str | None = None) -> dict:
    model_lower = model.lower()
    class_lower = (model_class_name or "").lower()
-    if "sana-video" in model_lower or "sana_video" in model_lower or "sanavideo" in class_lower:
-        return _MODEL_PRESETS["sana_720p" if "720p" in model_lower else "sana_480p"]
    if "lingbot" in model_lower or "lingbotvideo" in class_lower:
        return _MODEL_PRESETS["lingbot"]
    if "ltx" in class_lower or "ltx" in model_lower:
@@ -10,10 +10,12 @@ Generate text from text or image inputs with vLLM-Omni's shared offline entrypoi
 | --- | --- | --- | --- |
 | `ByteDance-Seed/BAGEL-7B-MoT` | Yes | Yes | `vllm_omni/deploy/bagel.yaml` |
 | `tencent/HunyuanImage-3.0-Instruct` | Yes | Yes | `vllm_omni/deploy/hunyuan_image3_ar.yaml` |
+| `bytedance-research/MammothModa2-Preview` | Yes | Yes | `vllm_omni/deploy/mammoth_moda2_ar.yaml` |
+| `bytedance-research/MammothModa2-Dev` | Yes | Yes | `vllm_omni/deploy/mammoth_moda2_ar.yaml` |

 The script recognizes these model families from `config.json`, applies the
-model-specific prompt format, and selects an AR-only deploy for HunyuanImage-3.
-BAGEL uses its registered default deploy.
+model-specific prompt format, and selects an AR-only deploy for HunyuanImage-3
+and MammothModa2. BAGEL uses its registered default deploy.

 ## Text-To-Text

@@ -33,6 +35,22 @@ python examples/offline_inference/x_to_text/x_to_text.py \
  --prompt "Explain multimodal inference in three concise sentences."
 ```

+### MammothModa2
+
+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model bytedance-research/MammothModa2-Preview \
+  --prompt "Explain multimodal inference in three concise sentences."
+```
+
+The same command supports the Dev checkpoint:
+
+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model bytedance-research/MammothModa2-Dev \
+  --prompt "Explain multimodal inference in three concise sentences."
+```
+
 ## Image-To-Text

 Pass one image with `--image`. The default question can be replaced with any
@@ -54,6 +72,16 @@ python examples/offline_inference/x_to_text/x_to_text.py \
  --prompt "Describe this image in detail."
 ```

+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model bytedance-research/MammothModa2-Preview \
+  --image image.png \
+  --prompt "Describe this image in detail."
+```
+
+Use `bytedance-research/MammothModa2-Dev` in the same command to run I2T with
+the Dev checkpoint.
+
 ## Key Arguments

 | Argument | Default | Description |
@@ -74,4 +102,4 @@ python examples/offline_inference/x_to_text/x_to_text.py \

 The committed HunyuanImage-3 AR-only default uses four GPUs with TP=4. To use a
 different topology, pass a custom AR-only YAML through `--deploy-config`.
-BAGEL uses its repository deploy default unless overridden.
+BAGEL and MammothModa2 use their repository deploy defaults unless overridden.
@@ -15,6 +15,7 @@ from vllm_omni.model_extras import build_x_to_text_prompt, get_x_to_text_model_f

 _REPO_ROOT = Path(__file__).resolve().parents[3]
 _HUNYUAN_AR_DEPLOY_CONFIG = _REPO_ROOT / "vllm_omni" / "deploy" / "hunyuan_image3_ar.yaml"
+_MAMMOTH_AR_DEPLOY_CONFIG = _REPO_ROOT / "vllm_omni" / "deploy" / "mammoth_moda2_ar.yaml"


 def parse_args() -> argparse.Namespace:
@@ -61,7 +62,7 @@ def main() -> None:
    omni_kwargs: dict[str, Any] = {
        "model": args.model,
        "mode": "image-to-text" if image is not None else "text-to-text",
-        "trust_remote_code": args.trust_remote_code or family == "hunyuan_image3",
+        "trust_remote_code": args.trust_remote_code or family in {"hunyuan_image3", "mammoth_moda2"},
        "log_stats": args.log_stats,
    }
    if args.enforce_eager is not None:
@@ -70,6 +71,8 @@ def main() -> None:
        omni_kwargs["deploy_config"] = args.deploy_config
    elif family == "hunyuan_image3":
        omni_kwargs["deploy_config"] = str(_HUNYUAN_AR_DEPLOY_CONFIG)
+    elif family == "mammoth_moda2":
+        omni_kwargs["deploy_config"] = str(_MAMMOTH_AR_DEPLOY_CONFIG)

    omni = Omni(**omni_kwargs)
    try:
@@ -0,0 +1,64 @@
+# AudioX online serving
+
+Launches the `AudioXPipeline` behind vLLM-Omni's OpenAI-compatible chat endpoint and provides a
+minimal Python client that covers all six tasks (`t2a`, `t2m`, `v2a`, `v2m`, `tv2a`, `tv2m`).
+
+## Start the server
+
+```bash
+cd examples/online_serving/audiox
+bash run_server.sh                 # defaults: MODEL=zhangj1an/AudioX, PORT=8099
+```
+
+Environment overrides: `MODEL`, `PORT`, `DIFFUSION_ATTENTION_BACKEND`.
+
+## Call from Python
+
+```bash
+# text-to-audio
+python openai_chat_client.py --task t2a \
+    --prompt "Fireworks burst twice, followed by a period of silence before a clock begins ticking." \
+    --output t2a.wav
+
+# text-to-music
+python openai_chat_client.py --task t2m \
+    --prompt "Uplifting ukulele tune for a travel vlog" \
+    --output t2m.wav
+
+# video-to-audio (no text)
+python openai_chat_client.py --task v2a --video path/to/clip.mp4 --output v2a.wav
+
+# text+video-to-audio
+python openai_chat_client.py --task tv2a \
+    --prompt "drum beating sound and human talking" \
+    --video path/to/clip.mp4 \
+    --output tv2a.wav
+```
+
+The client sends:
+
+- `num_inference_steps`, `guidance_scale`, `seed`, `audiox_task`, `seconds_start`,
+  `seconds_total`, `sigma_min`, `sigma_max` as flattened keys under `extra_body`. The server
+  routes the declared knobs (see `vllm_omni/model_extras/audiox.py`) into the pipeline's
+  `sampling_params.extra_args`.
+- For `v2*` / `tv2*` tasks, the video as a `video_url` content item (data URI for local files)
+
+## curl
+
+```bash
+curl -sS -X POST http://localhost:8099/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "zhangj1an/AudioX",
+    "messages": [{"role": "user", "content": [{"type": "text", "text": "Uplifting ukulele"}]}],
+    "extra_body": {
+      "num_inference_steps": 250,
+      "guidance_scale": 7.0,
+      "seed": 42,
+      "audiox_task": "t2m",
+      "seconds_total": 10.0,
+      "sigma_min": 0.3,
+      "sigma_max": 500.0
+    }
+  }' > t2m.json
+```
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""AudioX OpenAI-compatible chat client.
+
+AudioX supports 6 tasks (t2a, t2m, v2a, v2m, tv2a, tv2m). Text-only tasks send the prompt as the
+chat message; video-conditioned tasks additionally attach the video via a ``video_url`` content
+item (data URI for local files). Task + generation knobs (steps, cfg, sigma range, seconds, seed)
+are sent as flattened keys under ``extra_body``; the server routes the declared knobs into
+``sampling_params.extra_args`` (see ``vllm_omni/model_extras/audiox.py``).
+
+Usage:
+  python openai_chat_client.py --task t2a --prompt "Fireworks burst twice..." --output t2a.wav
+  python openai_chat_client.py --task tv2a --prompt "drum beating" --video clip.mp4 -o tv2a.wav
+"""
+
+from __future__ import annotations
+
+import argparse
+import base64
+import io
+import mimetypes
+import sys
+from pathlib import Path
+
+import requests
+import soundfile
+import torch
+
+VIDEO_TASKS = frozenset({"v2a", "v2m", "tv2a", "tv2m"})
+TEXT_TASKS = frozenset({"t2a", "t2m", "tv2a", "tv2m"})
+
+
+def _to_data_url(path: str) -> str:
+    mime, _ = mimetypes.guess_type(path)
+    mime = mime or "video/mp4"
+    with open(path, "rb") as f:
+        data = base64.b64encode(f.read()).decode("ascii")
+    return f"data:{mime};base64,{data}"
+
+
+def _save_wav(audio: torch.Tensor, path: Path, sample_rate: int) -> None:
+    audio = audio.to(torch.float32)
+    audio = audio / audio.abs().max().clamp(min=1e-8)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    # soundfile expects channels-last (T, C); project convention is (C, T).
+    soundfile.write(str(path), audio.clamp(-1.0, 1.0).cpu().T.numpy(), sample_rate, subtype="PCM_16")
+
+
+def _decode_audio_from_response(body: dict) -> tuple[torch.Tensor, int]:
+    for choice in body.get("choices", []):
+        audio_obj = choice.get("message", {}).get("audio")
+        if not (isinstance(audio_obj, dict) and audio_obj.get("data")):
+            continue
+        data, sr = soundfile.read(io.BytesIO(base64.b64decode(audio_obj["data"])), dtype="float32", always_2d=True)
+        return torch.from_numpy(data).transpose(0, 1), sr
+    brief = {k: v for k, v in body.items() if k != "choices"}
+    raise RuntimeError(f"no audio in response message.audio: {brief}")
+
+
+def main() -> int:
+    p = argparse.ArgumentParser(description="AudioX OpenAI chat client")
+    p.add_argument("--task", required=True, choices=["t2a", "t2m", "v2a", "v2m", "tv2a", "tv2m"])
+    p.add_argument("--prompt", "-p", default="", help="Text prompt (required for t2*/tv2*).")
+    p.add_argument("--video", help="Video path or URL (required for v2*/tv2*).")
+    p.add_argument("--output", "-o", default="audiox_out.wav")
+    p.add_argument("--server", "-s", default="http://localhost:8099")
+    p.add_argument("--model", default="zhangj1an/AudioX")
+    p.add_argument("--steps", type=int, default=250)
+    p.add_argument("--guidance-scale", type=float, default=7.0)
+    p.add_argument("--seed", type=int, default=42)
+    p.add_argument("--seconds-total", type=float, default=10.0)
+    p.add_argument("--seconds-start", type=float, default=0.0)
+    p.add_argument("--sigma-min", type=float, default=0.03)
+    p.add_argument("--sigma-max", type=float, default=1000.0)
+    args = p.parse_args()
+
+    if args.task in VIDEO_TASKS and not args.video:
+        print(f"ERROR: task {args.task!r} requires --video", file=sys.stderr)
+        return 2
+    if args.task in TEXT_TASKS and not args.prompt.strip() and args.task not in {"v2a", "v2m"}:
+        print(f"ERROR: task {args.task!r} requires --prompt", file=sys.stderr)
+        return 2
+
+    content: list[dict] = [{"type": "text", "text": args.prompt}]
+    if args.task in VIDEO_TASKS:
+        vurl = args.video if args.video.startswith(("http://", "https://")) else _to_data_url(args.video)
+        content.append({"type": "video_url", "video_url": {"url": vurl}})
+
+    payload = {
+        "model": args.model,
+        "messages": [{"role": "user", "content": content}],
+        "extra_body": {
+            "num_inference_steps": args.steps,
+            "guidance_scale": args.guidance_scale,
+            "seed": args.seed,
+            "audiox_task": args.task,
+            "seconds_start": args.seconds_start,
+            "seconds_total": args.seconds_total,
+            "sigma_min": args.sigma_min,
+            "sigma_max": args.sigma_max,
+        },
+    }
+
+    print(f"POST {args.server}/v1/chat/completions  task={args.task} steps={args.steps}")
+    r = requests.post(
+        f"{args.server}/v1/chat/completions",
+        headers={"Content-Type": "application/json"},
+        json=payload,
+        timeout=600,
+    )
+    r.raise_for_status()
+    audio, sr = _decode_audio_from_response(r.json())
+    _save_wav(audio, Path(args.output), sr)
+    dur = audio.shape[-1] / sr
+    print(f"saved {args.output}  sr={sr}Hz  duration={dur:.2f}s  channels={audio.shape[0]}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -0,0 +1,16 @@
+#!/bin/bash
+# AudioX online serving startup script.
+
+MODEL="${MODEL:-zhangj1an/AudioX}"
+PORT="${PORT:-8099}"
+DIFFUSION_ATTENTION_BACKEND="${DIFFUSION_ATTENTION_BACKEND:-FLASH_ATTN}"
+
+echo "Starting AudioX server..."
+echo "Model: $MODEL"
+echo "Port: $PORT"
+echo "Diffusion attention backend: $DIFFUSION_ATTENTION_BACKEND"
+
+DIFFUSION_ATTENTION_BACKEND="$DIFFUSION_ATTENTION_BACKEND" \
+    vllm serve "$MODEL" --omni \
+        --model-class-name AudioXPipeline \
+        --port "$PORT"
@@ -1,19 +1,5 @@
 # Image-To-Video

-This example demonstrates how to deploy image-to-video models for online
-video generation using vLLM-Omni.
-
-## Supported Models
-
-| Model | Model ID |
-|-------|----------|
-| Wan2.2 I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` |
-| SANA-Video 2B (480p) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` |
-| SANA-Video 2B (720p) | `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` |
-
-For SANA-Video 2B image-to-video online serving, see the
-[SANA-Video 2B recipe](../../../recipes/NVIDIA/SANA-Video-2B.md).
-
 This example demonstrates online image-to-video generation with vLLM-Omni.
 The startup script defaults to Wan2.2, but `MODEL` can select any supported
 image-to-video model. The existing curl helper remains a concrete Wan2.2
@@ -1,67 +0,0 @@
-#!/bin/bash
-# Submit and download a standard SANA-Video-2B image-to-video request.
-
-set -euo pipefail
-
-INPUT_IMAGE="${INPUT_IMAGE:-}"
-BASE_URL="${BASE_URL:-http://localhost:8099}"
-OUTPUT_PATH="${OUTPUT_PATH:-sana_video_i2v.mp4}"
-WIDTH="${WIDTH:-832}"
-HEIGHT="${HEIGHT:-480}"
-POLL_INTERVAL="${POLL_INTERVAL:-2}"
-
-if [ -z "$INPUT_IMAGE" ] || [ ! -f "$INPUT_IMAGE" ]; then
-    echo "Set INPUT_IMAGE to an existing image file."
-    exit 1
-fi
-
-create_response="$(
-    curl -sS -X POST "${BASE_URL}/v1/videos" \
-        -H "Accept: application/json" \
-        -F "prompt=A cat turns toward the camera with smooth, natural motion." \
-        -F "negative_prompt=blurry, low quality, temporal artifacts" \
-        -F "input_reference=@${INPUT_IMAGE}" \
-        -F "width=${WIDTH}" \
-        -F "height=${HEIGHT}" \
-        -F "num_frames=81" \
-        -F "fps=16" \
-        -F "num_inference_steps=50" \
-        -F "guidance_scale=6.0" \
-        -F "seed=42" \
-        -F 'extra_params={"motion_score":30}'
-)"
-
-video_id="$(echo "${create_response}" | jq -r '.id')"
-if [ -z "${video_id}" ] || [ "${video_id}" = "null" ]; then
-    echo "Failed to create SANA-Video job:"
-    echo "${create_response}" | jq .
-    exit 1
-fi
-
-while true; do
-    status_response="$(curl -sS "${BASE_URL}/v1/videos/${video_id}")"
-    status="$(echo "${status_response}" | jq -r '.status')"
-
-    case "${status}" in
-        queued|in_progress)
-            echo "Video job ${video_id} status: ${status}"
-            sleep "${POLL_INTERVAL}"
-            ;;
-        completed)
-            break
-            ;;
-        failed)
-            echo "SANA-Video generation failed:"
-            echo "${status_response}" | jq .
-            exit 1
-            ;;
-        *)
-            echo "Unexpected status response:"
-            echo "${status_response}" | jq .
-            exit 1
-            ;;
-    esac
-done
-
-curl -sS -L "${BASE_URL}/v1/videos/${video_id}/content" -o "${OUTPUT_PATH}"
-echo "Saved SANA-Video to ${OUTPUT_PATH}"
@@ -1,13 +0,0 @@
-#!/bin/bash
-# SANA-Video-2B native image-to-video serving.
-
-set -euo pipefail
-
-MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
-PORT="${PORT:-8099}"
-
-vllm serve "$MODEL" \
-    --omni \
-    --model-class-name SanaImageToVideoPipeline \
-    --dtype bfloat16 \
-    --port "$PORT"
@@ -1,18 +0,0 @@
-#!/bin/bash
-# SANA-Video-2B image-to-video serving through the Diffusers adapter.
-# Both 480p and 720p checkpoints are validated. For 720p:
-# MODEL=Efficient-Large-Model/SANA-Video_2B_720p_diffusers bash run_server_sana_video_diffusers.sh
-# INPUT_IMAGE=input.jpg WIDTH=1280 HEIGHT=704 bash run_curl_sana_video.sh
-
-set -euo pipefail
-
-MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
-PORT="${PORT:-8099}"
-
-vllm serve "$MODEL" \
-    --omni \
-    --model-class-name SanaImageToVideoPipeline \
-    --diffusion-load-format diffusers \
-    --diffusion-attention-backend TORCH_SDPA \
-    --dtype bfloat16 \
-    --port "$PORT"
@@ -25,6 +25,7 @@ For the full list of supported architectures across all modalities, see
 | Qwen3-TTS | `Qwen/Qwen3-TTS-12Hz-1.7B-{CustomVoice,VoiceDesign,Base}` | ✓ (Base) | ✓ (PCM + WebSocket) | ✓ (presets + `/v1/audio/voices` upload) | ✓ (standard + FastRTC) |
 | VoxCPM2 | `openbmb/VoxCPM2` | ✓ | ✓ (AudioWorklet via gradio) | — | ✓ |
 | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✓ (gated upstream) | ✓ | ✓ (presets) | ✓ |
+| SoulX-Singer | `Soul-AILab/SoulX-Singer` | ✓ (prompt audio) | — (batch only) | — (prompt + target audio) | — (chat client) |

 CosyVoice3 is intentionally absent: no online example exists for it yet. See its [offline section](../../offline_inference/text_to_speech/README.md#cosyvoice3) instead.

@@ -714,3 +715,65 @@ The demo handles voice-preset selection and reference-audio upload. `voxtral_tts
 - Voice presets are listed on the HF model card (`mistralai/Voxtral-4B-TTS-2603`).
 - Voice cloning is gated upstream and may require a recent `mistral_common`.
 - A standalone CLI client is not yet shipped; the gradio demo is the canonical reference for now.
+
+---
+
+## SoulX-Singer
+
+Singing voice synthesis (SVS) and conversion (SVC) at 24 kHz. Single-stage DiT with inline preprocess. Uses the `/v1/chat/completions` endpoint with multimodal input (`prompt_audio` + `target_audio`).
+
+### Prerequisites
+
+Download DiT and preprocess weights, then set up separate SVS / SVC view directories and install dependencies as described in the [offline README](../../offline_inference/text_to_speech/README.md#soulx-singer). `config.json` `architectures` field is the single source of truth for SVS vs SVC — point `MODEL` at the matching directory.
+
+### Launch
+
+```bash
+# SVS (default)
+export MODEL=/path/to/SoulX-Singer
+export PREPROCESS=/path/to/SoulX-Singer-Preprocess
+bash examples/online_serving/text_to_speech/soulxsinger/run_server.sh
+
+# SVC
+export MODE=svc
+export MODEL=/path/to/SoulX-Singer-svc
+bash examples/online_serving/text_to_speech/soulxsinger/run_server.sh
+```
+
+Or equivalently, set `SOULX_PREPROCESS_WEIGHTS_DIR` and launch directly:
+```bash
+export SOULX_PREPROCESS_WEIGHTS_DIR=$PREPROCESS
+vllm serve $MODEL --omni \
+    --deploy-config vllm_omni/deploy/soulxsinger_${MODE}.yaml \
+    --port 8192 --trust-remote-code --enforce-eager
+```
+
+### Sending requests
+
+Audio paths must be reachable from the server host (local filesystem or data URL). The client sends prompt vocal via `input_audio` and target accompaniment via `extra_args['target_audio']`.
+
+```bash
+# Default demo audio: tests/assets/soulxsinger/zh_prompt.mp3 + music.mp3
+python examples/online_serving/text_to_speech/soulxsinger/openai_chat_client.py \
+    --prompt-audio /path/on/server/zh_prompt.mp3 \
+    --target-audio /path/on/server/music.mp3 \
+    --preprocess-weights-dir /path/on/server/SoulX-Singer-Preprocess \
+    -o output.wav
+```
+
+Use precomputed metadata to skip online preprocess with following command:
+```bash
+python examples/online_serving/text_to_speech/soulxsinger/openai_chat_client.py \
+    --prompt-metadata-path /path/on/server/zh_prompt.json \
+    --target-metadata-path /path/on/server/music.json \
+    --audio-path /path/on/server/zh_prompt.mp3 \
+    -o output.wav
+```
+
+`SOULX_PREPROCESS_WEIGHTS_DIR` makes `--preprocess-weights-dir` optional. See `openai_chat_client.py --help` for `--vocal-sep`, `--language`, `--num-inference-steps`, `--guidance-scale`, and `--seed`.
+
+### Notes
+
+- Output: 24 kHz mono WAV; batch only.
+- Defaults match upstream: `--guidance-scale 3.0`, `--seed 42`, `--auto-shift` on.
+- SVS `--control`: `score` or `melody`. MIDI / lyric QC: upstream `midi_editor` only.
@@ -0,0 +1,186 @@
+#!/usr/bin/env python3
+"""SoulX-Singer OpenAI-compatible chat client (SVS / SVC).
+
+Sends prompt audio via ``input_audio`` and target accompaniment via
+``extra_args['target_audio']`` (server-local path). For integrated preprocess,
+also pass ``preprocess_weights_dir`` in ``extra_args``.
+
+Usage:
+  python openai_chat_client.py \\
+      --prompt-audio /path/on/server/zh_prompt.mp3 \\
+      --target-audio /path/on/server/music.mp3 \\
+      --preprocess-weights-dir /path/on/server/SoulX-Singer-Preprocess \\
+      -o output.wav
+"""
+
+from __future__ import annotations
+
+import argparse
+import base64
+import io
+import sys
+from pathlib import Path
+
+import requests
+import soundfile
+import torch
+
+
+def _audio_to_data_url(path: Path) -> str:
+    with path.open("rb") as handle:
+        data = base64.b64encode(handle.read()).decode("ascii")
+    return f"data:audio/mpeg;base64,{data}"
+
+
+def _save_wav(audio: torch.Tensor, path: Path, sample_rate: int) -> None:
+    audio = audio.to(torch.float32)
+    peak = audio.abs().max().clamp(min=1e-8)
+    audio = audio / peak
+    path.parent.mkdir(parents=True, exist_ok=True)
+    soundfile.write(str(path), audio.clamp(-1.0, 1.0).cpu().T.numpy(), sample_rate, subtype="PCM_16")
+
+
+def _decode_audio_from_response(body: dict) -> tuple[torch.Tensor, int]:
+    for choice in body.get("choices", []):
+        audio_obj = choice.get("message", {}).get("audio")
+        if isinstance(audio_obj, dict) and audio_obj.get("data"):
+            data, sr = soundfile.read(
+                io.BytesIO(base64.b64decode(audio_obj["data"])),
+                dtype="float32",
+                always_2d=True,
+            )
+            return torch.from_numpy(data).transpose(0, 1), sr
+    brief = {k: v for k, v in body.items() if k != "choices"}
+    raise RuntimeError(f"no audio in response message.audio: {brief}")
+
+
+def main() -> int:
+    repo_root = Path(__file__).resolve().parents[4]
+    default_assets = repo_root / "tests" / "assets" / "soulxsinger"
+
+    parser = argparse.ArgumentParser(description="SoulX-Singer online chat client")
+    parser.add_argument("--port", type=int, default=8192)
+    parser.add_argument("--model", default="Soul-AILab/SoulX-Singer")
+    parser.add_argument(
+        "--prompt-audio",
+        default=str(default_assets / "zh_prompt.mp3"),
+        help="Prompt vocal audio (path on server if using extra_args, or local for input_audio)",
+    )
+    parser.add_argument(
+        "--target-audio",
+        default=str(default_assets / "music.mp3"),
+        help="Target accompaniment path on the server (extra_args['target_audio'])",
+    )
+    parser.add_argument(
+        "--prompt-metadata-path",
+        default=None,
+        help="SVS precomputed prompt metadata.json",
+    )
+    parser.add_argument(
+        "--target-metadata-path",
+        default=None,
+        help="SVS precomputed target metadata.json",
+    )
+    parser.add_argument(
+        "--audio-path",
+        default=None,
+        help="SVS prompt vocal wav for precomputed metadata",
+    )
+    parser.add_argument("--preprocess-weights-dir", default=None)
+    parser.add_argument("--output", "-o", default="soulxsinger_out.wav")
+    parser.add_argument("--svc", action="store_true", help="Use SVC mode knobs")
+    parser.add_argument("--language", default="Mandarin")
+    parser.add_argument("--num-inference-steps", type=int, default=32)
+    parser.add_argument("--guidance-scale", type=float, default=3.0)
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="Optional CFM seed. Omit for non-deterministic sampling.",
+    )
+    parser.add_argument(
+        "--auto-shift",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Auto pitch shift (default: on, original upstream infer.sh)",
+    )
+    parser.add_argument(
+        "--control",
+        default="melody",
+        choices=["melody", "score"],
+        help="SVS control mode",
+    )
+    parser.add_argument("--vocal-sep", action="store_true")
+    args = parser.parse_args()
+
+    meta_paths = (args.prompt_metadata_path, args.target_metadata_path, args.audio_path)
+    if any(meta_paths) and not all(meta_paths):
+        print(
+            "ERROR: precomputed metadata requires --prompt-metadata-path, "
+            "--target-metadata-path, and --audio-path together.",
+            file=sys.stderr,
+        )
+        return 2
+
+    extra_args: dict = {
+        "vocal_sep": args.vocal_sep,
+        "auto_shift": args.auto_shift,
+        "pitch_shift": 0,
+    }
+    if all(meta_paths):
+        extra_args.update(
+            {
+                "prompt_metadata_path": str(Path(args.prompt_metadata_path).expanduser().resolve()),
+                "target_metadata_path": str(Path(args.target_metadata_path).expanduser().resolve()),
+                "audio_path": str(Path(args.audio_path).expanduser().resolve()),
+            }
+        )
+        content = [{"type": "text", "text": "soulx-singer"}]
+    else:
+        prompt_path = Path(args.prompt_audio).expanduser().resolve()
+        if not prompt_path.is_file():
+            print(f"ERROR: prompt audio not found: {prompt_path}", file=sys.stderr)
+            return 2
+        extra_args["prompt_audio"] = str(prompt_path)
+        extra_args["target_audio"] = str(Path(args.target_audio).expanduser().resolve())
+        if args.preprocess_weights_dir:
+            extra_args["preprocess_weights_dir"] = str(Path(args.preprocess_weights_dir).expanduser().resolve())
+        content = [
+            {"type": "text", "text": "soulx-singer"},
+            {
+                "type": "input_audio",
+                "input_audio": {"data": _audio_to_data_url(prompt_path), "format": "mp3"},
+            },
+        ]
+    if not args.svc:
+        extra_args["language"] = args.language
+        extra_args["control"] = args.control
+
+    payload = {
+        "model": args.model,
+        "modalities": ["audio"],
+        "messages": [{"role": "user", "content": content}],
+        "num_inference_steps": args.num_inference_steps,
+        "guidance_scale": args.guidance_scale,
+        "extra_args": extra_args,
+    }
+    if args.seed is not None:
+        payload["seed"] = args.seed
+
+    print(f"POST http://localhost:{args.port}/v1/chat/completions")
+    response = requests.post(
+        f"http://localhost:{args.port}/v1/chat/completions",
+        headers={"Content-Type": "application/json"},
+        json=payload,
+        timeout=1800,
+    )
+    response.raise_for_status()
+    audio, sample_rate = _decode_audio_from_response(response.json())
+    _save_wav(audio, Path(args.output), sample_rate)
+    duration = audio.shape[-1] / sample_rate
+    print(f"saved {args.output}  sr={sample_rate}Hz  duration={duration:.2f}s")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -0,0 +1,39 @@
+#!/bin/bash
+# Launch vLLM-Omni server for SoulX-Singer (single-stage DiT, preprocess inline).
+#
+# Usage:
+#   MODEL=/path/to/SoulX-Singer PREPROCESS=/path/to/Preprocess \
+#
+# Audio paths in client extra_args must be readable on the server host.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../../../.." && pwd)"
+
+MODEL="${MODEL:-Soul-AILab/SoulX-Singer}"
+MODE="${MODE:-svs}"
+PORT="${PORT:-8192}"
+GPUS="${GPUS:-0}"
+
+if [[ "$MODE" == "svc" ]]; then
+  DEPLOY_CONFIG="${DEPLOY_CONFIG:-$REPO_ROOT/vllm_omni/deploy/soulxsinger_svc.yaml}"
+else
+  DEPLOY_CONFIG="${DEPLOY_CONFIG:-$REPO_ROOT/vllm_omni/deploy/soulxsinger_svs.yaml}"
+fi
+
+echo "Starting SoulX-Singer server"
+echo "  MODEL=$MODEL"
+echo "  MODE=$MODE"
+echo "  PORT=$PORT"
+echo "  DEPLOY_CONFIG=$DEPLOY_CONFIG"
+echo "  CUDA_VISIBLE_DEVICES=$GPUS"
+
+CUDA_VISIBLE_DEVICES="$GPUS" \
+vllm serve "$MODEL" \
+    --omni \
+    --deploy-config "$DEPLOY_CONFIG" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --trust-remote-code \
+    --enforce-eager
@@ -11,11 +11,6 @@ This example demonstrates how to deploy text-to-video models for online video ge
 | Wan2.2 T2V | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` |
 | LTX-2 | `Lightricks/LTX-2` |
 | Helios (Base / Mid / Distilled) | `BestWishYsh/Helios-Base`, `Helios-Mid`, `Helios-Distilled` |
-| SANA-Video 2B (480p) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` |
-| SANA-Video 2B (720p) | `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` |
-
-For SANA-Video 2B text-to-video online serving, see the
-[SANA-Video 2B recipe](../../../recipes/NVIDIA/SANA-Video-2B.md).

 ## Wan2.2 T2V

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-BASE_URL="${BASE_URL:-http://localhost:8091}"
-OUTPUT_PATH="${OUTPUT_PATH:-sana_video.mp4}"
-
-curl -sS -X POST "$BASE_URL/v1/videos/sync" \
-    -F "prompt=A cinematic tracking shot of a sailboat crossing the ocean at sunset" \
-    -F "height=480" \
-    -F "width=832" \
-    -F "num_frames=81" \
-    -F "fps=16" \
-    -F "num_inference_steps=50" \
-    -F "guidance_scale=6.0" \
-    -F "seed=42" \
-    -F 'extra_params={"motion_score":30}' \
-    --output "$OUTPUT_PATH"
@@ -1,11 +0,0 @@
-#!/bin/bash
-# SANA-Video-2B text-to-video serving.
-
-MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
-PORT="${PORT:-8091}"
-
-vllm serve "$MODEL" \
-    --omni \
-    --model-class-name SanaVideoPipeline \
-    --dtype bfloat16 \
-    --port "$PORT"
@@ -1,12 +0,0 @@
-#!/bin/bash
-# SANA-Video through the black-box Diffusers backend adapter.
-
-MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
-PORT="${PORT:-8091}"
-
-vllm serve "$MODEL" \
-    --omni \
-    --diffusion-load-format diffusers \
-    --diffusion-attention-backend TORCH_SDPA \
-    --dtype bfloat16 \
-    --port "$PORT"
@@ -73,8 +73,6 @@ dev = [
    "FlagEmbedding==1.4.0", # for hunyuan_image3 tests
    "kernels==0.14.1", # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
    "s3tokenizer==0.3.0", # for tts test
-    # Provides top-level cosyvoice2; do not co-install stepaudio2-minicpmo,
-    # whose top-level stepaudio2 package conflicts with this distribution.
    "step-audio2==1.0.0", # Step-Audio2 / MiniCPM-o 4.5 Token2Wav (flashcosyvoice / hyperpyyaml)
 ]

@@ -137,6 +135,13 @@ docs = [
    "pydantic",
 ]

+soulx-svs = [
+    "funasr>=1.0.0",
+    "g2pM>=0.1.2.5",
+    "g2p-en>=2.1.0",
+    "ToJyutping>=3.2.0",
+]
+
 [project.urls]
 Homepage = "https://github.com/vllm-project/vllm-omni"
 Repository = "https://github.com/vllm-project/vllm-omni"
@@ -187,7 +192,6 @@ select = [
    "N",      # pep8-naming
    "UP",     # pyupgrade
    "TID251", # flake8-tidy-imports.banned-api
-    "TC004",  # flake8-type-checking: runtime-used imports
 ]
 ignore = [
    "E203",  # whitespace before ':' (conflicts with black)
@@ -268,7 +272,6 @@ markers = [
    # specified computation resources marks (auto-added)
    # Tag with [hardware-resource] for tests.helpers.mark.get_hardware_mark_list().
    "H100: [hardware-resource] Tests that require H100 GPU",
-    "H800: [hardware-resource] Tests that require H800 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",
@@ -1,113 +0,0 @@
-# HiDream-O1-Image
-
-> HiDream-O1 text-to-image generation through the shared offline example
-
-## Summary
-
- Vendor: HiDream.ai
- Model: `HiDream-ai/HiDream-O1-Image`
- Task: Text-to-image
- Mode: Offline inference
- Maintainer: Community
-
-## When to use this recipe
-
-Use this recipe to generate native-resolution images with HiDream-O1 on one
-or two CUDA GPUs. The two-GPU command shards the Qwen3-VL attention and MLP
-weights, as well as the timestep and pixel projections.
-
-## References
-
- Upstream model:
-  [`HiDream-ai/HiDream-O1-Image`](https://huggingface.co/HiDream-ai/HiDream-O1-Image)
- Related offline example:
-  [`examples/offline_inference/text_to_image/text_to_image.py`](../../examples/offline_inference/text_to_image/text_to_image.py)
- Related issue:
-  [`vllm-project/vllm-omni#3479`](https://github.com/vllm-project/vllm-omni/issues/3479)
-
-## Hardware Support
-
-## GPU
-
-### 1x H100 80GB
-
-#### Environment
-
- OS: Linux
- Python: Match the repository requirements for your checkout
- Driver / runtime: NVIDIA CUDA
- vLLM version: Match the repository requirements for your checkout
- vLLM-Omni version or commit: Use the commit you are deploying from
-
-#### Command
-
-```bash
-python examples/offline_inference/text_to_image/text_to_image.py \
-  --model HiDream-ai/HiDream-O1-Image \
-  --prompt "A cat is sitting next to a sign that says 'HiDream-O1 vLLM-Omni'" \
-  --height 2048 \
-  --width 2048 \
-  --num-inference-steps 50 \
-  --guidance-scale 5.0 \
-  --seed 42 \
-  --output hidream_o1_output.png
-```
-
-#### Verification
-
-```bash
-python -c "from PIL import Image; im = Image.open('hidream_o1_output.png'); print(im.mode, im.size)"
-```
-
-Expected output:
-
-```text
-RGB (2048, 2048)
-```
-
-#### Notes
-
- The pipeline supports BF16 text-to-image generation.
- Requested sizes are matched to the closest resolution supported by the checkpoint.
-
-### 2x H100 80GB
-
-#### Environment
-
- OS: Linux
- Python: Match the repository requirements for your checkout
- Driver / runtime: NVIDIA CUDA
- vLLM version: Match the repository requirements for your checkout
- vLLM-Omni version or commit: Use the commit you are deploying from
-
-#### Command
-
-```bash
-python examples/offline_inference/text_to_image/text_to_image.py \
-  --model HiDream-ai/HiDream-O1-Image \
-  --tensor-parallel-size 2 \
-  --prompt "A cat is sitting next to a sign that says 'HiDream-O1 vLLM-Omni'" \
-  --height 2048 \
-  --width 2048 \
-  --num-inference-steps 50 \
-  --guidance-scale 5.0 \
-  --seed 42 \
-  --output hidream_o1_tp2_output.png
-```
-
-#### Verification
-
-```bash
-python -c "from PIL import Image; im = Image.open('hidream_o1_tp2_output.png'); print(im.mode, im.size)"
-```
-
-Expected output:
-
-```text
-RGB (2048, 2048)
-```
-
-#### Notes
-
- Tensor parallelism shards the attention, MLP, timestep, and pixel projection weights across both GPUs.
- Image editing and image-conditioned generation are not supported.
@@ -0,0 +1,212 @@
+# MammothModa2
+
+> MammothModa2-Preview and MammothModa2-Dev unified understanding and generation
+
+## Summary
+
+- Vendor: ByteDance Research
+- Models: `bytedance-research/MammothModa2-Preview`, `bytedance-research/MammothModa2-Dev`
+- Tasks: Preview and Dev text-to-image (AR → DiT); Dev text/image understanding
+- Mode: Offline inference
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe to run MammothModa2-Preview through shared task-oriented
+offline examples. Text-to-image uses the shared image example (`text_to_image.py`) instead of a model-specific script.
+The generic example formats the AR prompt, drives the AR → DiT stage pipeline,
+and forwards MammothModa2-specific generation parameters through the
+pipeline-declared `extra_body` contract.
+
+MammothModa2's DiT stage consumes its inputs through the multi-stage kwargs
+interface (not `OmniDiffusionRequest`), so its generation knobs
+(`text_guidance_scale`, `cfg_range`, `num_inference_steps`) are passed via
+`--extra-body` rather than the standard `--num-inference-steps` / `--cfg-scale`
+flags. Image size uses the standard `--height` / `--width` flags.
+
+## References
+
+- Upstream model:
+  [`bytedance-research/MammothModa2-Preview`](https://huggingface.co/bytedance-research/MammothModa2-Preview)
+- Dev model:
+  [`bytedance-research/MammothModa2-Dev`](https://huggingface.co/bytedance-research/MammothModa2-Dev)
+- Related offline example:
+  [`examples/offline_inference/text_to_image/text_to_image.py`](../../examples/offline_inference/text_to_image/text_to_image.py)
+- Related T2T/I2T example:
+  [`examples/offline_inference/x_to_text/x_to_text.py`](../../examples/offline_inference/x_to_text/x_to_text.py)
+- Declared parameters:
+  [`vllm_omni/model_extras/mammothmodal2_preview.py`](../../vllm_omni/model_extras/mammothmodal2_preview.py)
+- Deploy config:
+  [`vllm_omni/deploy/mammoth_moda2.yaml`](../../vllm_omni/deploy/mammoth_moda2.yaml)
+
+## Hardware Support
+
+The default deploy config runs both the AR and DiT stages on a single GPU
+(`devices: "0"`). The committed `gpu_memory_utilization` split (stage-0 AR `0.5`,
+stage-1 DiT `0.3`) is sized for an ~80 GB GPU. The model also fits on a 48 GB GPU
+after rebalancing the split so the AR weights (~23 GB) leave room for the KV
+cache — see the note under *1x L40S 48GB*.
+
+## GPU
+
+### 1x L40S 48GB
+
+> **48 GB config adjustment:** the committed
+> `vllm_omni/deploy/mammoth_moda2.yaml` uses
+> `gpu_memory_utilization` 0.5 / 0.3 (sized for ~80 GB). To fit on a 48 GB L40S,
+> set the stage-0 (AR) value to `0.8` and the stage-1 (DiT) value to `0.16`
+> before running. (On an ~80 GB GPU, leave the defaults unchanged.)
+
+#### Environment
+
+- OS: Linux
+- Python: Match the repository requirements for your checkout
+- Driver / runtime: NVIDIA CUDA environment with one L40S 48 GB (verified) or an
+  ~80 GB GPU for the default config
+- vLLM version: Match the repository requirements for your checkout
+- vLLM-Omni version or commit: Use the commit you are deploying from
+
+#### Offline Commands
+
+Download the model:
+
+```bash
+hf download bytedance-research/MammothModa2-Preview --local-dir ./MammothModa2-Preview
+```
+
+Run text-to-image with the shared offline example from the repository root. The
+deploy config sets `trust_remote_code`, so no extra flag is needed. Forward the
+MammothModa2 generation parameters as a JSON object through `--extra-body`:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model ./MammothModa2-Preview \
+  --deploy-config vllm_omni/deploy/mammoth_moda2.yaml \
+  --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
+  --height 1024 \
+  --width 1024 \
+  --extra-body '{"text_guidance_scale": 4.0, "cfg_range": [0.0, 1.0], "num_inference_steps": 50}' \
+  --output mammoth_t2i.png
+```
+
+The `--extra-body` JSON forwards MammothModa2-specific parameters into
+`OmniDiffusionSamplingParams.extra_args`. Keys are filtered against the model's
+declared `extra_body_params` (see
+[`vllm_omni/model_extras/mammothmodal2_preview.py`](../../vllm_omni/model_extras/mammothmodal2_preview.py)),
+so unknown keys for MammothModa2 are silently dropped:
+
+- `text_guidance_scale` — classifier-free guidance scale for the DiT stage
+  (default `9.0`; CFG is active only when `> 1.0`).
+- `cfg_range` — relative step range `[start, end]` over which CFG is applied
+  (default `[0.0, 1.0]`).
+- `num_inference_steps` — number of DiT denoising steps (default `50`).
+
+`--height` and `--width` must be multiples of 16.
+
+Run text-to-text through the shared understanding example. It recognizes the
+MammothModa2 checkpoint and automatically selects `mammoth_moda2_ar.yaml`:
+
+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model ./MammothModa2-Preview \
+  --prompt "Explain multimodal generation in three sentences."
+```
+
+Add an image for image-to-text or image summarization. The shared example
+uses MammothModa2's chat and vision-token template:
+
+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model ./MammothModa2-Preview \
+  --image /path/to/input.jpg \
+  --prompt "Please summarize the content of this image."
+```
+
+#### Verification
+
+The example writes the generated image to the `--output` path. Confirm the file
+exists and is a valid image:
+
+```bash
+ls -lh mammoth_t2i.png
+python -c "from PIL import Image; print(Image.open('mammoth_t2i.png').size)"
+```
+
+### 1x AMD MI300X, MammothModa2 Preview
+
+#### Environment
+
+- OS: Linux 6.8.0-134-generic, x86_64
+- Container: official ROCm image built from `docker/Dockerfile.rocm`
+- Python: 3.12.13
+- PyTorch: 2.11.0+gitd0c8b1f
+- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
+- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
+- vLLM version: 0.27.0+rocm723
+- vLLM Omni version or commit: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
+- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
+
+#### Offline Commands
+
+The checked run used the committed stage split, with `gpu_memory_utilization` set to 0.5 for AR and 0.3 for DiT:
+
+```bash
+python3 examples/offline_inference/text_to_image/text_to_image.py \
+    --model bytedance-research/MammothModa2-Preview \
+    --deploy-config vllm_omni/deploy/mammoth_moda2.yaml \
+    --prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
+    --height 1024 \
+    --width 1024 \
+    --seed 42 \
+    --extra-body '{"text_guidance_scale": 4.0, "cfg_range": [0.0, 1.0], "num_inference_steps": 50}' \
+    --enable-diffusion-pipeline-profiler \
+    --log-stats \
+    --output mammoth_t2i.png
+```
+
+#### Verification
+
+The first request took 85.224 seconds. The AR stage generated 4,161 visual tokens in 72.996 seconds, and the DiT stage took 12.163 seconds. AR weight loading used 21.4 GiB and took 8.250 seconds. DiT weight loading used 5.49 GiB and took 1.824 seconds. The largest one second whole device memory sample was 106.57 GiB, including the AR KV cache reserved by the 0.5 memory setting.
+
+The output was a valid 1024 by 1024 RGB PNG.
+
+## MammothModa2-Dev unified inference
+
+MammothModa2-Dev uses a Qwen3-VL AR backbone, while MammothModa2-Preview uses
+Qwen2.5-VL. vLLM-Omni selects the matching implementation from the nested
+`llm_config.model_type`; no checkpoint edits or `trust_remote_code` flag are
+required.
+
+Text-to-text and image-to-text use the AR-only deploy. Text-to-image loads the
+Qwen3 generation experts (`gen_mlp`), extra visual vocabulary and image head,
+then sends the generated visual tokens and hidden states to the DiT stage.
+
+Download the checkpoint:
+
+```bash
+hf download bytedance-research/MammothModa2-Dev --local-dir ./MammothModa2-Dev
+```
+
+Run text-to-text through the shared understanding example. It recognizes the
+Dev checkpoint as MammothModa2 and automatically selects
+`mammoth_moda2_ar.yaml`:
+
+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model ./MammothModa2-Dev \
+  --prompt "Explain multimodal generation in three sentences."
+```
+
+Add an image for image-to-text or image summarization:
+
+```bash
+python examples/offline_inference/x_to_text/x_to_text.py \
+  --model ./MammothModa2-Dev \
+  --image ./image.png \
+  --prompt "Please summarize the content of this image."
+```
+
+The Dev checkpoint is approximately 47.55 GiB on disk. In the verified AR-only
+run, loaded model weights used approximately 16.97 GiB of GPU memory before KV
+and encoder caches. Allow additional GPU memory for those caches and the input
+image.
@@ -110,10 +110,9 @@ outputs of each pair agreed to within 115 bytes.
 Four-GPU T2VA repeats: 273.8 s and 269.5 s, peaking at 15,560 MiB and
 15,612 MiB. Four-GPU Ref2VA is a single successful end-to-end run (545.2 s,
 16,448 MiB). Follow-up Ref2VA requests on this build completed diffusion but
-then hit the engine's async output wait during post-compute D2H, so they are
-not reported as repeats. That wait was a hardcoded 30 s when these numbers were
-taken; it is now `VLLM_OMNI_ASYNC_OUTPUT_TIMEOUT` (default 600 s), so a rerun on
-current main should not lose these repeats to the bound.
+then hit the engine's hardcoded 30 s async output wait
+(`_ASYNC_OUTPUT_TIMEOUT` in `diffusion_engine.py`) during post-compute D2H,
+so they are not reported as repeats.

 Relative to the two-GPU baseline on the same shape, four-GPU TP2×USP2 was about
 1.6× faster for both tasks, while per-GPU peak HBM stayed in the mid-15 GiB
@@ -118,21 +118,6 @@ On Atlas 800I A3 (64 GB HBM per device) the combined service does not fit at
 above) or HSDP — see
 [§ Memory and attention optimizations](#memory-and-attention-optimizations-a3).

-### CPU MP4 response encoding (Atlas A2)
-
-Non-streaming MP4 responses use one public automatic encoder. It checks the
-runtime frame shape, common dtype, and RGB channel-plane contiguity for every
-request; compatible inputs use direct planar PyAV frames, while unsupported
-inputs fall back to the legacy muxer before opening the PyAV container. No CLI
-flag, model declaration, or user configuration is required. Streaming fMP4
-output remains on its existing incremental path.
-
-Current performance and correctness validation is limited to one Atlas A2 host
-with 8x Ascend 910B4-1 NPUs, the `FL2VA`/`t2va` partition, one request at a
-time, 1344x768 at 24 fps, and 5, 8.7, and 15 second requests. This optimization
-only changes CPU MP4 response encoding; it does not change DiT execution or
-stage 0.
-
 ### Optional optimizations

 Two independent optimizations may be enabled on top of the configuration
@@ -65,31 +65,6 @@ On AMD ROCm, install without the `[fa4]` extra (FA4 is CUDA-only) and use
 `ffmpeg` and `ffprobe` must be available on `PATH`. They are used for
 reference-video preparation and MP4 output.

-## CPU MP4 response encoding
-
-For CUDA and ROCm deployments, non-streaming MP4 responses are encoded on the
-host CPU through PyAV/libx264 after generation. The response encoder selects the
-path automatically at runtime: inputs with a supported frame shape, common dtype,
-and RGB per-channel contiguous layout use direct planar frames; other inputs
-fall back to the legacy path before the PyAV container is opened. No CLI flag,
-model declaration, or user configuration is required. Streaming fMP4 output is
-unchanged.
-
-A community benchmark on 2x Xeon 8480C reported the following comparison
-between the legacy and direct planar paths
-([full result](https://github.com/vllm-project/vllm-omni/pull/6288#issuecomment-5337546499)):
-
-| Metric | Legacy | Direct planar | Change |
-| --- | ---: | ---: | ---: |
-| Median wall time | 1.805 s | 1.394 s | -22.8% |
-| Median process CPU time | 3.613 s | 3.207 s | -11.2% |
-| Peak RSS | 3182 MiB | 2794 MiB | -387 MiB (-12.2%) |
-
-Across the 1.0-8.7 s sweep, wall-time improvement was approximately 21.8-22.6%;
-outputs were byte-identical and full decode passed. This is evidence for host
-CPU response encoding. Actual gains depend on the CPU and runtime, and should
-not be interpreted as GPU, DiT, or stage 0 speedups.
-
 ## Start a server

 Pass the repository ID directly. The pipeline uses `FL2VA` for model discovery
@@ -183,7 +158,7 @@ vllm serve "${MODEL}" \
 Use the profile that matches the per-GPU memory capacity:

 | Profile | GPUs | Starting shape | Resident DiT blocks | Attention | Execution | Status |
-| --- | ---: | ---: | ---: | --- | --- | --- |
+|---|---:|---:|---:|---|---|---|
 | `rtx5090` | 2 x 32 GB | 1344x768 | 20 | cuDNN attention | eager | Target-hardware validated |
 | `rtx4090` | 2 x 24 GB | 1024x576 | 12 | cuDNN attention | eager | Capacity-proxy starting point |

@@ -202,8 +177,8 @@ increasing it on a different request shape.
 At vLLM-Omni commit `ae6577ea`, one full 50-step T2VA request completed on
 2 x RTX 5090 without OOM:

-| Shape    | Frames        | Client E2E | Sampled peak/GPU       | Output validation                                            |
-| -------: | ------------: | ---------: | ---------------------: | -----------------------------------------------------------: |
+| Shape | Frames | Client E2E | Sampled peak/GPU | Output validation |
+|---:|---:|---:|---:|---|
 | 1344x768 | 124 at 24 FPS | 8 min 38 s | approximately 22.6 GiB | H.264 video + 32 kHz stereo AAC; full `ffmpeg` decode passed |

 This is a single end-to-end validation run, not a warmed multi-run latency
@@ -272,7 +247,7 @@ vllm serve "${MODEL}" \
 Do not add `--enforce-eager` to this performance configuration. The first
 request includes regional compilation; warm the server once before measuring
 steady-state latency. H3 is CFG-distilled, so `--cfg-parallel-size` must remain
-`1`. The H3 VAE supports its native `tile` mode, not
+1. The H3 VAE supports its native `tile` mode, not
 `spatial_shard_height` or `spatial_shard_width`.

 ### Attention Backends
@@ -380,63 +355,6 @@ No restart is needed: `task=fl2va` routes to `FL2VA/transformer`, while
 `task=ref2va` routes to
 `Ref2VA/transformer`. T2VA uses the FL2VA DiT.

-### Step execution and continuous batching
-
-H3 implements the step-wise execution contract, so the scheduler can admit and
-retire requests between denoise steps instead of running one request end to
-end. Add the feature gate, then raise `--max-num-seqs` to co-batch:
-
-```bash
--step-execution --max-num-seqs 4
-```
-
-Co-batched requests are packed into a single sequence that keeps one attention
-document per request, so a batch costs one DiT forward. That packing requires
-`--diffusion-attention-backend FLASH_ATTN`; other backends fall back to one
-forward per request. `--max-num-seqs 1` keeps the conservative single-request
-step path. Cache acceleration (`--cache-backend`) is not available in step mode.
-
-!!! warning "Co-batching does not improve H3 throughput for large simultaneous requests"
-    Keep `--max-num-seqs 1` unless you specifically need scheduler-level control
-    (admitting and retiring requests between denoise steps). Measured on two
-    H100s (TP2, BF16, 672x384, 209 frames, 30 steps, 4 requests at concurrency
-    4 with `--request-rate inf`, i.e. all four submitted at time 0; one packed
-    request is 16384 rows):
-
-    | Configuration | Wall time | Mean latency | Peak memory |
-    |---------------|-----------|--------------|-------------|
-    | request mode | 174.8 s | 111.5 s | 72.4 GB |
-    | `--step-execution --max-num-seqs 1` | 179.0 s | 113.8 s | 72.4 GB |
-    | `--step-execution --max-num-seqs 4` | 182.1 s | 175.7 s | 78.3 GB |
-
-    A single H3 denoise step is a compute-bound dense GEMM over an already long
-    packed sequence, so fusing N requests costs N times the FLOPs and buys almost
-    no amortization — unlike LLM decoding, which is memory-bandwidth bound.
-    Going from one request per step to four cuts the per-request denoise cost
-    only from 1.323 s to 1.291 s (2.4%), which the step-mode bookkeeping then
-    spends. Mean latency degrades further because co-batched requests finish
-    together instead of staggered. Quantization moves the absolute numbers
-    without changing this: with online `int8` the same workload runs in 153.3 s
-    at 56.9 GB (request mode), and `--max-num-seqs 4` is still 5.0% slower than
-    request mode.
-
-    Two workloads outside this table are unmeasured and may behave differently:
-
-    - **Staggered arrivals (admission latency).** A single 16384-row H3 request
-      already saturates the GPUs, so submitting all requests at time 0 is the
-      one arrival pattern where co-batching cannot win: it can only bunch
-      completions. In request mode a new request queues behind the whole
-      in-flight generation (~45 s at this size); step mode admits at the next
-      denoise-step boundary (~1.3 s). Whether that shows up as a wall-clock
-      benefit under Poisson arrivals is not yet measured for H3. To reproduce,
-      run request mode vs `--step-execution --max-num-seqs 4` with
-      `diffusion_benchmark_serving.py --request-rate 0.05` (roughly one request
-      every 20 s) and report mean / p95 latency, which then includes queueing.
-    - **Small requests.** The numbers above are drawn from 672x384 / 209-frame
-      requests. A short clip at lower resolution packs a few thousand rows and
-      may not saturate the hardware; co-batching may amortize better there.
-      This is also unmeasured.
-
 ### Online FP8 quantization

 MiniMax H3 supports online FP8 quantization of both the DiT and the Qwen3-VL
@@ -587,7 +505,7 @@ vLLM-Omni with MiniMax H3 support, BF16. gfx942 rows measured with the
 `0.26.0+rocm723` wheel (HIP 7.2).

 | Workload | Configuration | Observed result |
-| ---------- | --------------- | ----------------- |
+|----------|---------------|-----------------|
 | T2VA, 1344x768, 209 frames, 50 steps | 4x gfx942 (MI300X), FLASH_ATTN, USP4, text-enc TP4, VAE PP4 tile | encode 0.09 s, denoise 244.04 s, decode 4.15 s, 267.42 s client E2E; H.264 24 FPS + 32 kHz stereo AAC |
 | FL2VA, 1344x768, 209 frames, 50 steps | 4x gfx942 (MI300X), FLASH_ATTN, USP4, text-enc TP4, VAE PP4 tile | encode 13.98 s, denoise 257.58 s, decode 4.11 s, 287.07 s client E2E; H.264 24 FPS + 32 kHz stereo AAC |
 | T2VA, 832x480, ~4 s, 40 steps | 1x gfx950 (MI350), FLASH_ATTN, CPU offload | valid MP4 (H.264 + synced audio); ~0.73 s/denoise-step (~1.37 it/s), ~55 s client E2E incl. warmup |
@@ -769,7 +687,7 @@ at most 15 seconds combined.
 ## Official input matrix and limits

 | Task | Supported references | Limits |
-| ------ | ---------------------- | -------- |
+|------|----------------------|--------|
 | T2VA | text only | prompt must be non-empty |
 | FL2VA | first image, last image, or ordered first+last images | at most 2 images; `frame_indices` is `[0]`, `[-1]`, or `[0,-1]` |
 | Ref2VA | image-only, image+image, image+video, video+audio, and mixed image/video/audio | images ≤9, videos ≤3, audios ≤3, total references ≤12; audio requires a visual reference |
@@ -800,17 +718,13 @@ Omitting `quality` preserves the startup default: it uses the reference path
 normally, or the server-configured profile when the server was started with
 `--cache-backend cache_dit`.

-Turbo is independent of this quality switch: `quality` selects a Cache-DiT
-policy, while Turbo changes the active LoRA weights and sampling schedule. See
-[LoRA](#lora) below.
-
 The following result was measured on 4× NVIDIA H200 with SP4, text-encoder
 TP4, 1344×768, 124 frames, 24 FPS, and 50 inference steps. One full
 `lossless` warmup was excluded, followed by three fixed prompt/seed pairs in
 balanced switch order.

 | `quality` | Median inference latency | Speedup | SSIM vs `lossless` | PSNR vs `lossless` | Expected trade-off |
-| --- | ---: | ---: | ---: | ---: | --- |
+|---|---:|---:|---:|---:|---|
 | `lossless` | 85.49 s | 1.00× | 1.0000 | exact | Native reference path |
 | `high` | 63.36 s | 1.35× | 0.9709 | 34.98 dB | Faster with measured same-seed deviation |

@@ -820,49 +734,10 @@ balanced switch order.
 > workload. The values above apply to this deployment and are not universal
 > guarantees. `lossless` remains the exact reference path.

-## LoRA
-
-### Turbo LoRA
-
-Only the native Diffusers 4-step FL2VA/T2VA v1.0 artifact is supported:
-
-```text
-lightx2v/Minimax-h3-Turbo/minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
-```
-
-Download only that file:
-
-```bash
-export TURBO_DIR=/path/to/minimax-h3-turbo
-export TURBO_FILE=minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
-hf download lightx2v/Minimax-h3-Turbo "${TURBO_FILE}" --local-dir "${TURBO_DIR}"
-export TURBO_LORA="${TURBO_DIR}/${TURBO_FILE}"
-```
-
-Start from a non-offloaded FL2VA server command and add
-`--task-type fl2va --lora-backend peft --lora-path "${TURBO_LORA}"`.
-`--lora-path` preloads the adapter; each request still activates it and uses
-the published sampling settings:
-
-```bash
-F 'num_inference_steps=5' \
-F 'flow_shift=6' \
-F 'extra_params={"task":"t2va","duration":4.4,"audio_flow_shift":3.0}' \
-F "lora={\"name\":\"h3-turbo-v1.0\",\"path\":\"${TURBO_LORA}\",\"scale\":1.0}"
-```
-
-For FL2VA, change `task` and add `input_reference` as shown above. The 8-step,
-ComfyUI, Ref2VA, and v1.1 artifacts are not supported. This integration is
-dynamic-only and does not support prefusion, DLO, or LoRA composition.
-It also rejects model-level, layerwise, and distributed layerwise offload;
-the legacy dynamic LoRA tensors do not participate in those weight lifecycles.
-The five requested sigma points produce the four denoiser evaluations expected
-by the Turbo artifact.
-
 ## Key parameters

 | Parameter | Recommended value | Notes |
-| ----------- | ------------------- | ------- |
+|-----------|-------------------|-------|
 | `quality` | omitted or `lossless` | Request-level quality intent; `high` dynamically installs H3's conservative Cache-DiT profile |
 | `extra_params.force_refresh_step_hint` | omitted | Optional positive 1-based denoising-step hint for an active Cache-DiT request; pair with `extra_params.force_refresh_step_policy`=`once` or `repeat` |
 | `task` | `t2va`, `fl2va`, or `ref2va` | Passed in `extra_params`; selects the task-specific DiT |
@@ -887,9 +762,9 @@ Users can also use a ComfyUI frontend to interact with a hosted MiniMax-H3 servi
 The four-GPU recommendation was measured on four NVIDIA B300 GPUs with one
 excluded warmup followed by three requests.

-| Workload                               | Configuration                               | Observed result                      |
-| -------------------------------------- | ------------------------------------------- | ------------------------------------ |
-| FL2VA, 209 frames, 1248x768            | no offload, U4, VPP4 tile, regional compile | 86.964 s mean HTTP client latency    |
+| Workload | Configuration | Observed result |
+|----------|---------------|-----------------|
+| FL2VA, 209 frames, 1248x768 | no offload, U4, VPP4 tile, regional compile | 86.964 s mean HTTP client latency |
 | Two-video Ref2VA, 362 frames, 1344x768 | no offload, U4, VPP4 tile, regional compile | 784.394 s accounted model-stage mean |

 These measurements describe the validated shapes rather than a general
@@ -981,12 +856,7 @@ vllm serve "${MODEL_ROOT}/FL2VA" \
 - TeaCache is calibrated for FL2VA only; Ref2VA requests run uncached.
 - Combined serving requires sibling `FL2VA` and `Ref2VA` directories, loads
  both task-specific DiTs, and loads shared components once from `FL2VA`.
- Request mode executes one generation request per diffusion batch. Use
-  `--step-execution` with `--max-num-seqs N` to admit several requests at once
-  (see
-  [Execution modes](https://github.com/vllm-project/vllm-omni/blob/main/docs/user_guide/diffusion/execution_modes.md));
-  step
-  mode does not support `cache_backend`.
+- H3 currently executes one generation request per diffusion batch.
 - The first regional-compile request is a warmup and should not be included in
  steady-state performance measurements.
 - The serving path accepts fewer references than the model supports. H3 documents up
@@ -1,245 +0,0 @@
-# SANA-Video 2B
-
-> Native text-to-video and image-to-video generation at 480p and 720p,
-> plus a validated Diffusers-adapter compatibility path
-
-## Summary
-
- Vendor: NVIDIA
- Model: `Efficient-Large-Model/SANA-Video_2B_480p_diffusers`,
-  `Efficient-Large-Model/SANA-Video_2B_720p_diffusers`
- Task: Text-to-video and image-to-video
- Mode: Offline inference and OpenAI-compatible online serving
- Maintainer: Community
-
-## When to use this recipe
-
-Use the native `SanaVideoPipeline` for T2V and
-`SanaImageToVideoPipeline` for I2V. Both native pipelines support the 480p
-and 720p checkpoints. Use `--diffusion-load-format diffusers` when you need
-the black-box Diffusers compatibility baseline; adapter T2V and I2V are
-validated at both resolutions.
-
-The native pipeline loads the 480p checkpoint through
-`DistributedAutoencoderKLWan` and the 720p checkpoint through
-`DistributedAutoencoderKLLTX2Video`. These are vLLM-Omni distributed wrappers
-around the corresponding Diffusers autoencoders, not independent VAE
-implementations. The denoising loop also intentionally loads Diffusers'
-`DPMSolverMultistepScheduler` from the checkpoint to preserve its scheduler
-configuration and numerical behavior.
-
-## References
-
- Upstream project: <https://github.com/NVlabs/Sana>
- Model cards:
-  [480p](https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_480p_diffusers),
-  [720p](https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_720p_diffusers)
- Diffusers documentation: <https://huggingface.co/docs/diffusers/api/pipelines/sana_video>
- Online serving guides:
-  [Text-to-Video](../../docs/user_guide/examples/online_serving/text_to_video.md),
-  [Image-to-Video](../../docs/user_guide/examples/online_serving/image_to_video.md)
- Offline T2V example:
-  [`examples/offline_inference/text_to_video/text_to_video.py`](../../examples/offline_inference/text_to_video/text_to_video.py)
- Offline I2V example:
-  [`examples/offline_inference/image_to_video/image_to_video.py`](../../examples/offline_inference/image_to_video/image_to_video.py)
- Support discussion: [vLLM-Omni issue #5432](https://github.com/vllm-project/vllm-omni/issues/5432)
-
-## Hardware Support
-
-## GPU
-
-### 1x RTX 5090 32GB
-
-#### Environment
-
- OS: Ubuntu 22.04.5 LTS
- Python: 3.12.3
- Driver / runtime: NVIDIA driver 580.95.05; PyTorch 2.11.0+cu130
- Diffusers: 0.38.0
- vLLM version: 0.26.0
- vLLM-Omni version or commit: PR #5508, commit `22037901`
-
-#### Command
-
-##### Native text-to-video inference
-
-```bash
-python examples/offline_inference/text_to_video/text_to_video.py \
-  --model Efficient-Large-Model/SANA-Video_2B_720p_diffusers \
-  --model-class-name SanaVideoPipeline \
-  --prompt "A cat walking on the grass, facing the camera." \
-  --negative-prompt "blurry, low quality, temporal artifacts" \
-  --height 704 --width 1280 --num-frames 81 \
-  --num-inference-steps 50 --guidance-scale 6 \
-  --extra-body '{"motion_score": 30}' \
-  --fps 16 --seed 42 --output sana_video_720p.mp4
-```
-
-For 480p, select `SANA-Video_2B_480p_diffusers` and use
-`--height 480 --width 832`.
-
-##### Native image-to-video inference
-
-SANA checkpoints declare `SanaVideoPipeline` in `model_index.json`, so I2V
-must be selected explicitly with `--model-class-name
-SanaImageToVideoPipeline`.
-
-```bash
-python examples/offline_inference/image_to_video/image_to_video.py \
-  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  --model-class-name SanaImageToVideoPipeline \
-  --image input.jpg \
-  --prompt "A cat turns toward the camera with smooth, natural motion." \
-  --negative-prompt "blurry, low quality, temporal artifacts" \
-  --height 480 --width 832 --num-frames 81 \
-  --num-inference-steps 50 --guidance-scale 6 \
-  --extra-body '{"motion_score": 30}' \
-  --fps 16 --seed 42 --output sana_video_i2v_480p.mp4
-```
-
-The same pipeline class supports the 720p checkpoint through vLLM-Omni's
-distributed LTX-2 VAE wrapper; use `--height 704 --width 1280`.
-
-For online I2V serving:
-
-```bash
-MODEL=Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  bash examples/online_serving/image_to_video/run_server_sana_video.sh
-
-INPUT_IMAGE=input.jpg OUTPUT_PATH=sana_video_i2v.mp4 \
-  bash examples/online_serving/image_to_video/run_curl_sana_video.sh
-```
-
-##### Native online serving
-
-```bash
-MODEL=Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  bash examples/online_serving/text_to_video/run_server_sana_video.sh
-
-bash examples/online_serving/text_to_video/run_curl_sana_video.sh
-```
-
-To run the black-box compatibility backend for T2V, replace the server script
-with `run_server_sana_video_diffusers.sh`. The same `/v1/videos` request
-works; `num_frames` is adapted to Diffusers' `frames` argument. The script
-selects `TORCH_SDPA` because SANA-Video uses an attention mask that the
-AITER-backed Diffusers attention path does not accept.
-
-##### Diffusers-adapter image-to-video serving
-
-The validated I2V adapter commands are:
-
-```bash
-# 480p
-MODEL=Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
-  bash examples/online_serving/image_to_video/run_server_sana_video_diffusers.sh
-
-INPUT_IMAGE=input.jpg OUTPUT_PATH=sana_video_i2v_adapter.mp4 \
-  bash examples/online_serving/image_to_video/run_curl_sana_video.sh
-
-# 720p
-MODEL=Efficient-Large-Model/SANA-Video_2B_720p_diffusers \
-  bash examples/online_serving/image_to_video/run_server_sana_video_diffusers.sh
-
-INPUT_IMAGE=input.jpg WIDTH=1280 HEIGHT=704 \
-  OUTPUT_PATH=sana_video_i2v_adapter_720p.mp4 \
-  bash examples/online_serving/image_to_video/run_curl_sana_video.sh
-```
-
-#### Verification
-
-Check the encoded 720p output metadata after running a generation command:
-
-```bash
-ffprobe -v error -select_streams v:0 \
-  -show_entries stream=width,height,r_frame_rate,nb_frames \
-  -of default=noprint_wrappers=1 sana_video_720p.mp4
-```
-
-The standard 720p request above should report:
-
-```text
-width=1280
-height=704
-r_frame_rate=16/1
-nb_frames=81
-```
-
-For 480p, expect `width=832` and `height=480`.
-
-The automated serving matrix covers both checkpoint variants:
-
-| Backend | 480p T2V | 720p T2V | 480p I2V | 720p I2V |
-| --- | --- | --- | --- | --- |
-| Native vLLM-Omni | Validated | Validated | Validated | Validated |
-| Diffusers adapter | Validated | Validated | Validated | Validated |
-
-Use the native `SanaVideoPipeline` and `SanaImageToVideoPipeline` for the
-primary SANA execution paths. The Diffusers adapter is retained as a
-validated compatibility/reference backend.
-
-#### Notes
-
- Key flags: select I2V explicitly with `--model-class-name
-  SanaImageToVideoPipeline`; SANA checkpoint `model_index.json` files declare
-  the T2V class. Pass `motion_score` through `--extra-body`. The supplied
-  Diffusers-adapter server scripts select `TORCH_SDPA` because the AITER-backed
-  path does not accept SANA's attention mask.
- Output profile: 81 frames at 16 FPS is the standard checkpoint profile and
-  produces approximately five seconds of video. Minute-scale generation
-  requires the separate LongSANA/LongLive block-autoregressive workflow.
- Backend boundary: the native pipelines and Transformer are owned by
-  vLLM-Omni. The 480p Wan VAE and 720p LTX-2 VAE run through vLLM-Omni
-  distributed wrappers derived from the corresponding Diffusers VAE classes.
-  The denoising loop intentionally retains the checkpoint-compatible
-  Diffusers `DPMSolverMultistepScheduler`.
- Known limitations:
-    - Sequence/tensor/CFG parallelism, Cache-DiT, TeaCache, and step execution
-    are not validated for the native pipeline.
-    - The Diffusers backend is a compatibility path and does not provide native
-    vLLM-Omni parallelism or continuous batching.
-    - Native describes pipeline and Transformer ownership, not a zero-Diffusers
-    dependency guarantee.
-
-### 1x NVIDIA L20X
-
-#### Environment
-
- Driver: 570.133.20
- PyTorch: 2.13.0+cu130
- Diffusers: 0.38.0
- vLLM: 0.27.0
- Nsight Systems: 2026.1
- vLLM-Omni commit: `54fcc6d`
-
-#### Same-workload T2V comparison
-
-Both backends used
-`Efficient-Large-Model/SANA-Video_2B_480p_diffusers`, the same T2V prompt
-and seeds, 832×480 output, 81 frames, and four inference steps. One warmup
-request was excluded, followed by three measured requests.
-
-| Metric | Native vLLM-Omni | Diffusers adapter | Native delta |
-| --- | ---: | ---: | ---: |
-| E2E latency | 7.1600 ± 0.0043 s | 6.9761 ± 0.0517 s | +2.636% |
-| Throughput | 0.1397 req/s | 0.1433 req/s | -2.568% |
-| Peak reserved VRAM | 23,936 MiB | 19,060 MiB | +25.582% |
-| Initialization | 11.669 s | 20.236 s | -42.336% |
-
-The peak-memory result is not a Transformer-only comparison. For the 480p
-checkpoint, the native pipeline deliberately loads and decodes with the Wan
-VAE in FP32. This matches the upstream SANA-Video reference and the recorded
-accuracy-golden setup for this integration. The adapter launch uses
-`--dtype bfloat16` as the Diffusers pipeline-wide dtype, so its Wan VAE also
-runs in BF16.
-
-Peak VRAM is the CUDA allocator's maximum reserved memory over the complete
-request. It includes VAE decode activations and allocator-reserved blocks, not
-only live Transformer tensors. The different VAE precision policies are a
-material code-level difference and a likely contributor to the higher native
-high-water mark; the available full-request measurements do not isolate how
-much of the delta they account for. The 25.582% result should therefore not be
-interpreted as an isolated native Transformer memory regression. Exact
-attribution requires both paths to use the same VAE dtype. Changing the native
-480p VAE to BF16 would instead change the recorded golden/reference setup and
-requires a new accuracy study.
@@ -39,6 +39,7 @@ recipes/

 | Recipe | Task | Hardware |
 |--------|------|----------|
+| [`audiox/AudioX.md`](./audiox/AudioX.md) | Offline + online unified text/video→audio diffusion | 1x L4 24GB |
 | [`Baidu/ERNIE-Image.md`](./Baidu/ERNIE-Image.md) | Text-to-image online serving (ERNIE-Image 8B) | 1x or 2x RTX 4090 24GB |
 | [`Bagel/BAGEL-7B-MoT.md`](./Bagel/BAGEL-7B-MoT.md) | Text-to-image with shared online/offline examples | 1x A100 80GB / 2x CUDA GPUs |
 | [`Boogu/Boogu-Image.md`](./Boogu/Boogu-Image.md) | Text-to-image online serving (Boogu-Image-0.1-Base) | 1x A100/H100 40GB+ |
@@ -46,7 +47,6 @@ recipes/
 | [`ByteDance/Lance.md`](./ByteDance/Lance.md) | Unified AR+diffusion: text/img/video gen + understanding (Lance 3B) | 1x B300 / A100 80GB |
 | [`fishaudio/Fish-Speech-S2-Pro.md`](./fishaudio/Fish-Speech-S2-Pro.md) | Online serving for TTS | 1x A800 80GB |
 | [`Helios/Helios.md`](./Helios/Helios.md) | Text-to-video, image-to-video, and video-to-video generation | 1x NVIDIA H20 |
-| [`HiDream-ai/HiDream-O1-Image.md`](./HiDream-ai/HiDream-O1-Image.md) | Text-to-image with shared offline example | 1x or 2x H100 80GB |
 | [`inclusionAI/Ming-flash-omni-2.0.md`](./inclusionAI/Ming-flash-omni-2.0.md) | Online serving for multimodal chat + standalone TTS | 4x H100 / 1x H100 80GB |
 | [`inclusionAI/Ming-omni-tts.md`](./inclusionAI/Ming-omni-tts.md) | Offline + online dense Ming TTS/audio generation | 1x H100 80GB / 1x AMD MI300X (ROCm 7.2) |
 | [`k2-fsa/OmniVoice.md`](./k2-fsa/OmniVoice.md) | Offline and online multilingual text-to-speech | 1x AMD MI300X (ROCm 7.2) |
@@ -58,7 +58,6 @@ recipes/
 | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | LTX-2/LTX-2.3 text-to-video and image-to-video with synchronized audio | H200 141GB / 96GB-class GPU |
 | [`LTX/LTX-2.5.md`](./LTX/LTX-2.5.md) | LTX-2.5-Diffusers: Full/SFT one-stage and distilled two-stage T2V/I2V with synchronized audio | NVIDIA B300; cuDNN-qualified |
 | [`MammothModa2/MammothModa2.md`](./MammothModa2/MammothModa2.md) | Preview and Dev text-to-image (AR → DiT); Dev text/image understanding | Preview: 1x L40S 48GB / 1x ≥40GB GPU; Dev: 1x NVIDIA GPU with sufficient cache headroom |
-| [`NVIDIA/SANA-Video-2B.md`](./NVIDIA/SANA-Video-2B.md) | Native and Diffusers-adapter 480p/720p T2V+I2V | 1x RTX 5090 32GB |
 | [`meituan-longcat/LongCat-Video-Avatar-1.5.md`](./meituan-longcat/LongCat-Video-Avatar-1.5.md) | Audio-driven avatar video generation (AT2V / AI2V, single- and multi-speaker, AVC continuation) | 1x H100 80GB |
 | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
 | [`Robbyant/LingBot-Video.md`](./Robbyant/LingBot-Video.md) | Native dense and MoE text-to-video serving | Dense: 1x L20X; MoE: 1x L20X (~67.7 GiB peak) |
@@ -0,0 +1,130 @@
+# AudioX
+
+> AudioX MMDiT for unified audio + music generation: t2a / t2m / v2a / v2m / tv2a / tv2m
+
+## Summary
+
+- Vendor: HKUSTAudio (project), `zhangj1an/AudioX` weight bundle
+- Model: `zhangj1an/AudioX`
+- Task: Text/video → audio or music. Six tasks: `t2a`, `t2m`, `v2a`, `v2m`, `tv2a`, `tv2m`.
+- Mode: Offline inference + online serving (pure diffusion)
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe to run AudioX for sound-effect (`*2a`) or music (`*2m`) generation
+from a text prompt and/or video clip. AudioX is a unified diffusion transformer
+that produces stereo 44.1 kHz audio up to ~10 s per call.
+
+## References
+
+- Project page: <https://zeyuet.github.io/AudioX/>
+- vLLM-Omni weight bundle: <https://huggingface.co/zhangj1an/AudioX>
+- Pipeline: `vllm_omni.diffusion.models.audiox.pipeline_audiox.AudioXPipeline`
+- Input transforms: `vllm_omni.transformers_utils.processors.audiox`
+- Param contract: `vllm_omni/model_extras/audiox.py` (declared `extra_body` knobs)
+- Offline example: [`examples/offline_inference/text_to_audio/text_to_audio.py`](../../examples/offline_inference/text_to_audio/text_to_audio.py)
+- Online: standard OpenAI chat-completions endpoint (see commands below)
+
+## Hardware Support
+
+## GPU
+
+### 1x L4 24GB
+
+#### Environment
+
+- OS: Ubuntu 22.04
+- Python: 3.10+
+- Driver / runtime: CUDA 12.4
+- vLLM version: 0.20.0
+- vLLM-Omni version: 0.1.x
+
+#### Command
+
+AudioX uses the **standard** `text_to_audio` example offline and the standard
+OpenAI chat-completions endpoint online. All model-specific knobs
+(`audiox_task`, `seconds_start`, `seconds_total`, `sigma_min`, `sigma_max`,
+`cfg_rescale`, `video_path`, `audio_path`) are declared in
+`vllm_omni/model_extras/audiox.py` and routed through `extra_body` (online) /
+`--extra-body` (offline). The offline example also exposes `--task`, `--video`,
+`--audio-start`, and `--audio-length` shortcuts for the most common knobs.
+
+Offline — text tasks (`t2a` / `t2m`):
+
+```bash
+huggingface-cli download zhangj1an/AudioX --local-dir ./audiox_weights
+python examples/offline_inference/text_to_audio/text_to_audio.py \
+  --model ./audiox_weights --task t2a \
+  --prompt "Fireworks burst twice, followed by a clock ticking." \
+  --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0 \
+  --extra-body '{"sigma_min": 0.03, "sigma_max": 1000.0}' \
+  --output t2a.wav
+```
+
+Offline — video-conditioned tasks (`v2a` / `v2m` / `tv2a` / `tv2m`) need `--video`:
+
+```bash
+python examples/offline_inference/text_to_audio/text_to_audio.py \
+  --model ./audiox_weights --task tv2a \
+  --prompt "drum beating sound and human talking" \
+  --video https://zeyuet.github.io/AudioX/static/samples/V2M/1XeBotOFqHA.mp4 \
+  --num-inference-steps 250 --guidance-scale 6.0 --audio-length 10.0 \
+  --output tv2a.wav
+```
+
+Online:
+
+```bash
+DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN \
+  vllm serve zhangj1an/AudioX --omni --model-class-name AudioXPipeline --port 8099
+
+curl -sS -X POST http://localhost:8099/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "zhangj1an/AudioX",
+    "messages": [{"role": "user", "content": [{"type": "text", "text": "Uplifting ukulele"}]}],
+    "extra_body": {
+      "num_inference_steps": 250,
+      "guidance_scale": 7.0,
+      "seed": 42,
+      "audiox_task": "t2m",
+      "seconds_total": 10.0,
+      "sigma_min": 0.3,
+      "sigma_max": 500.0
+    }
+  }' > t2m.json
+```
+
+For `v2*` / `tv2*` online, attach the video as a `video_url` content item
+(data URI for local files) in the message `content`.
+
+#### Verification
+
+```bash
+# Health check
+curl http://localhost:8099/health
+
+# Listen to the saved file (stereo, 44.1 kHz, sigma_min=0.03, sigma_max=1000 — upstream defaults)
+ffprobe t2a.wav
+```
+
+#### Notes
+
+- Memory usage: ~10 GB peak with `num_inference_steps=250`, 10 s of audio.
+- Output rate: 44.1 kHz stereo regardless of requested rate.
+- Supported tasks: `t2a`, `t2m`, `v2a`, `v2m`, `tv2a`, `tv2m`. Pass via the
+  `audiox_task` knob (in `extra_body` online, or `--task` offline).
+- Video conditioning: `v2*` and `tv2*` require a video; offline via `--video`
+  (→ `video_path` knob), online via a `video_url` content item.
+- Cache acceleration is **not** supported (AudioXPipeline is in `_NO_CACHE_ACCELERATION`).
+- Tensor parallelism is supported via `--tensor-parallel-size` (DiT QKV is sharded with
+  `QKVParallelLinear`); cross-attention K/V is also TP-sharded.
+
+### Known limitations
+
+- Inference uses an inlined DPM-Solver++(3M) SDE sampler (k-diffusion port). Replacing it with
+  diffusers' `EDMDPMSolverMultistepScheduler` introduces a fixed ~861 Hz resonance and is not
+  recommended.
+- Generation is fixed at 10 s (configured by the bundle's `sample_size`); longer outputs require
+  a different bundle.
@@ -12,7 +12,7 @@
 transformers >= 5.10.1, < 5.15
 av>=14.0.0
 omegaconf>=2.3.0
-diffusers==0.40.0
+diffusers==0.38.0
 safetensors>=0.8.0
 accelerate==1.12.0
 soundfile>=0.13.1
@@ -0,0 +1,16 @@
+[
+  {
+    "index": "vocal_240_51240",
+    "language": "Mandarin",
+    "time": [
+      240,
+      51240
+    ],
+    "duration": "0.21 0.18 0.34 0.38 0.26 0.22 0.50 0.20 0.33 0.13 0.60 0.22 0.38 0.52 0.30 0.12 1.82 0.98 0.20 0.26 0.38 0.36 0.58 0.54 0.26 0.50 1.46 3.03 0.24 0.24 0.29 0.25 0.20 0.14 0.80 0.22 0.54 0.30 0.24 0.16 0.58 0.28 0.38 1.73 0.79 0.24 0.30 0.34 0.30 0.30 0.34 0.21 0.21 0.36 0.34 0.23 1.85 1.90 0.23 0.39 0.68 0.50 0.31 0.43 0.76 0.38 2.00 1.87 0.68 0.72 0.56 0.62 0.80 0.40 0.42 1.68 1.79 0.70 0.66 0.54 0.24 0.48 0.68 0.40 2.34 0.14",
+    "text": "<SP> 只 是 因 为 为 在 人 群 群 中 多 看 了 你 一 眼 <SP> 再 也 没 能 忘 掉 你 容 颜 <SP> 梦 想 着 着 偶 偶 然 然 有 <SP> 一 一 天 再 相 见 <SP> 从 此 我 开 始 始 孤 孤 单 思 念 念 <SP> 想 想 你 时 你 你 在 天 边 <SP> 想 你 时 你 在 眼 前 前 <SP> 想 你 时 你 你 在 脑 海 <SP>",
+    "phoneme": "<SP> zh_zhi3 zh_shi4 zh_yin1 zh_wei4 zh_wei2 zh_zai4 zh_ren2 zh_qun2 zh_qun2 zh_zhong1 zh_duo1 zh_kan4 zh_le5 zh_ni3 zh_yi1 zh_yan3 <SP> zh_zai4 zh_ye3 zh_mei2 zh_neng2 zh_wang4 zh_diao4 zh_ni3 zh_rong2 zh_yan2 <SP> zh_meng4 zh_xiang3 zh_zhe5 zh_zhe5 zh_ou3 zh_ou3 zh_ran2 zh_ran2 zh_you3 <SP> zh_yi1 zh_yi1 zh_tian1 zh_zai4 zh_xiang1 zh_jian4 <SP> zh_cong2 zh_ci3 zh_wo3 zh_kai1 zh_shi3 zh_shi3 zh_gu1 zh_gu1 zh_dan1 zh_si1 zh_nian4 zh_nian4 <SP> zh_xiang3 zh_xiang3 zh_ni3 zh_shi2 zh_ni3 zh_ni3 zh_zai4 zh_tian1 zh_bian1 <SP> zh_xiang3 zh_ni3 zh_shi2 zh_ni3 zh_zai4 zh_yan3 zh_qian2 zh_qian2 <SP> zh_xiang3 zh_ni3 zh_shi2 zh_ni3 zh_ni3 zh_zai4 zh_nao3 zh_hai3 <SP>",
+    "note_pitch": "0 64 64 64 66 68 66 66 66 64 64 64 64 66 64 60 61 0 63 63 63 64 66 63 61 59 56 0 68 66 66 68 68 66 66 64 64 0 64 66 61 61 61 64 0 62 63 63 64 64 66 65 66 61 58 59 56 0 69 71 66 68 68 71 66 64 61 0 66 61 68 66 64 63 61 59 0 71 66 68 68 71 66 64 61 0",
+    "note_type": "1 2 2 2 2 3 2 2 2 3 2 2 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 1 2 2 2 3 2 3 2 3 2 1 2 3 2 2 2 2 1 2 2 2 2 2 3 2 3 2 2 2 3 1 2 3 2 2 2 3 2 2 2 1 2 2 2 2 2 2 2 3 1 2 2 2 2 3 2 2 2 1",
+    "f0": "0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 318.9 332.2 331.2 323.9 312.3 0.0 0.0 0.0 0.0 0.0 0.0 340.9 342.3 340.2 337.4 333.5 329.3 327.8 328.3 329.1 330.5 331.9 331.5 329.2 327.6 327.4 327.7 329.6 330.5 329.5 328.1 328.5 330.1 330.4 329.5 328.9 331.0 333.6 332.8 331.7 330.2 330.6 330.9 329.5 327.6 332.4 343.5 363.9 368.7 369.3 368.2 366.7 371.3 382.0 392.9 402.3 409.8 413.8 416.3 416.4 415.4 414.1 413.4 415.3 417.4 418.3 413.6 416.0 0.0 0.0 0.0 0.0 368.2 362.5 361.3 362.0 362.9 364.3 366.4 367.5 368.4 368.9 368.1 367.6 369.7 371.1 371.1 369.5 368.0 367.3 366.6 367.1 369.5 372.0 371.3 368.2 368.1 369.6 369.9 373.4 376.5 360.2 285.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 384.0 376.7 372.0 371.7 373.9 376.3 375.5 361.2 333.0 317.8 319.0 329.1 333.5 333.2 333.2 333.3 328.0 314.5 0.0 0.0 0.0 320.6 326.5 332.9 335.7 334.1 329.9 327.4 326.3 326.6 329.1 330.5 330.0 328.0 326.7 328.6 331.1 332.4 332.1 332.8 333.2 333.1 331.8 328.6 318.9 293.2 323.6 327.5 330.0 332.8 331.1 319.4 272.2 0.0 0.0 0.0 0.0 278.5 294.5 310.0 318.5 322.7 327.4 334.1 337.2 336.3 328.5 321.0 324.2 341.5 362.2 374.7 375.8 370.5 366.0 365.2 368.7 371.7 374.4 373.0 368.7 370.4 374.8 375.1 372.5 368.7 363.0 357.1 359.0 366.8 377.1 379.5 371.5 359.9 351.0 358.4 371.7 377.9 375.8 367.8 359.7 363.6 375.3 379.9 377.5 372.9 359.8 345.2 337.1 334.5 333.4 332.8 329.6 319.2 292.8 261.7 253.4 262.8 273.0 278.1 279.3 278.6 277.6 277.9 278.3 278.0 277.3 276.7 275.6 275.3 276.1 277.8 278.2 277.7 277.6 277.2 276.4 275.5 273.9 271.8 270.7 274.7 282.7 285.5 281.4 273.1 264.4 262.2 268.9 278.4 284.3 283.6 277.3 268.5 262.2 263.2 270.7 278.8 285.0 287.3 282.0 272.1 265.2 266.0 272.1 278.2 283.8 285.1 281.4 272.6 267.2 269.1 276.1 282.7 289.3 290.5 285.6 273.9 267.5 267.4 271.8 277.7 282.6 283.1 277.8 269.8 265.6 268.6 273.8 281.1 286.2 287.2 282.3 270.4 264.5 263.3 263.0 268.5 268.3 268.7 270.9 277.2 277.4 284.3 285.7 282.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 311.5 306.4 306.0 306.8 305.8 305.0 307.2 310.8 312.4 313.5 314.4 314.6 313.7 312.2 311.1 311.1 312.8 314.6 312.7 306.5 299.7 302.7 308.7 315.2 314.2 310.9 309.4 309.8 310.6 310.7 310.8 311.4 312.2 313.0 312.5 312.2 312.2 311.9 312.0 312.2 311.4 306.1 303.3 309.3 315.0 325.3 324.4 323.9 325.0 326.4 327.7 327.7 328.1 328.6 327.8 328.3 328.3 327.6 328.6 336.1 347.2 359.4 371.0 374.9 371.3 366.5 369.5 375.5 376.2 370.2 365.7 367.3 372.9 376.9 375.3 368.4 360.1 358.0 365.4 380.8 383.5 380.4 374.4 367.6 365.9 372.0 376.3 377.9 375.7 372.6 368.6 361.2 353.7 0.0 309.4 302.3 299.2 300.6 305.7 308.5 308.6 309.0 310.9 310.9 309.5 308.6 309.0 311.3 313.0 314.4 314.5 313.0 312.1 311.2 308.0 299.5 295.8 295.0 285.7 278.0 280.1 281.2 279.9 278.3 278.3 279.5 279.7 272.2 259.1 0.0 0.0 0.0 240.7 236.0 233.8 233.5 234.6 235.7 235.6 236.2 239.8 245.4 248.8 250.1 250.4 249.9 249.1 247.2 241.7 231.7 216.0 197.6 192.1 197.2 204.3 206.5 205.4 201.7 200.8 203.5 208.2 210.9 210.3 207.1 203.5 202.3 202.1 202.8 205.1 207.3 208.1 207.3 205.4 202.1 199.3 200.5 206.1 212.9 212.9 206.5 197.1 191.1 191.2 196.8 203.6 207.8 208.7 206.4 201.7 197.3 197.4 200.7 206.5 209.9 209.9 206.9 203.3 201.2 203.4 207.7 210.5 208.9 207.6 206.5 204.2 202.0 203.6 205.7 210.8 213.1 214.3 210.6 204.1 199.7 202.3 211.9 217.7 215.1 215.0 215.3 213.3 0.0 0.0 216.3 0.0 0.0 195.2 209.0 205.3 201.3 196.8 195.7 195.8 195.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 431.9 425.8 421.0 417.7 419.3 422.7 423.4 421.1 417.9 411.9 408.6 411.5 414.9 419.6 426.7 424.2 0.0 0.0 0.0 280.8 279.4 369.7 371.8 370.6 369.8 370.5 370.5 371.4 374.6 374.0 361.0 332.7 328.0 0.0 359.0 365.0 371.1 373.8 371.9 369.7 369.6 371.0 378.3 390.4 403.4 414.1 417.9 417.5 417.0 416.6 415.7 414.9 414.5 413.8 413.5 413.6 413.8 415.2 416.9 417.3 415.4 415.0 414.1 411.4 409.2 403.5 392.7 375.6 367.5 365.8 368.8 370.7 370.8 369.0 367.0 366.0 366.2 367.8 369.2 368.8 367.2 366.7 367.3 367.3 368.8 368.5 366.8 365.3 364.0 363.0 365.0 367.3 367.7 365.5 364.0 364.9 367.4 370.3 371.3 369.9 366.5 364.1 368.3 385.8 406.7 409.4 399.7 376.0 357.2 359.9 372.4 377.5 366.9 345.1 320.6 315.8 321.8 329.0 329.9 326.4 324.5 324.8 324.9 325.3 325.1 324.7 325.9 326.8 326.6 326.6 326.4 325.4 323.2 322.5 323.9 326.2 328.6 329.9 329.5 329.0 328.2 327.7 327.9 329.6 331.2 331.4 334.2 340.1 335.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.Line truncated
+  }
+]
@@ -0,0 +1,16 @@
+[
+  {
+    "index": "vocal_320_10687",
+    "language": "Mandarin",
+    "time": [
+      320,
+      10687
+    ],
+    "duration": "0.23 0.34 0.26 0.70 0.52 0.46 0.36 0.44 0.14 0.24 0.64 0.47 0.51 1.10 0.28 0.38 0.32 0.32 0.38 0.32 0.31 0.19 1.45",
+    "text": "<SP> 除 了 想 你 你 <SP> 除 了 了 爱 你 你 <SP> 我 什 么 什 么 都 愿 愿 意",
+    "phoneme": "<SP> zh_chu2 zh_le5 zh_xiang3 zh_ni3 zh_ni3 <SP> zh_chu2 zh_le5 zh_le5 zh_ai4 zh_ni3 zh_ni3 <SP> zh_wo3 zh_shen2 zh_me5 zh_shen2 zh_me5 zh_dou1 zh_yuan4 zh_yuan4 zh_yi4",
+    "note_pitch": "0 62 65 67 67 69 0 67 69 67 65 67 69 67 67 66 64 64 60 60 65 67 0",
+    "note_type": "1 2 2 2 2 3 1 2 2 3 2 2 3 1 2 2 2 2 2 2 2 3 2",
+    "f0": "0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 294.1 288.2 290.6 294.6 295.7 292.8 291.5 294.4 295.4 294.7 293.5 292.2 294.4 295.8 293.2 297.7 320.1 338.1 348.3 348.7 344.4 342.6 346.2 354.8 356.9 353.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 401.1 403.8 405.0 402.1 398.4 395.6 393.3 392.0 391.4 390.3 390.2 390.3 390.3 391.3 392.8 393.6 391.7 390.5 391.4 391.6 391.5 393.5 393.8 390.8 387.4 387.8 389.3 390.8 392.2 391.6 390.2 389.8 389.1 388.4 390.0 395.5 397.2 396.7 395.5 395.1 394.6 394.9 395.6 395.2 394.6 395.4 395.9 394.0 391.7 390.7 391.7 392.6 391.6 395.7 405.8 441.7 462.5 463.8 450.2 430.9 414.5 415.2 426.7 439.8 454.4 462.9 447.8 422.5 400.6 403.2 423.5 451.3 482.3 492.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 435.9 414.3 406.2 402.8 398.2 396.7 396.7 395.2 398.3 416.9 441.6 446.9 441.5 437.3 435.3 432.9 431.6 432.3 434.8 436.7 433.8 422.8 406.7 390.9 382.0 381.7 384.7 382.7 368.1 357.4 355.6 355.1 352.4 348.1 346.4 348.7 351.6 355.0 354.4 351.7 349.9 349.2 348.1 346.0 345.4 344.2 344.4 345.5 346.6 349.0 349.7 349.1 349.5 349.6 349.7 349.4 349.5 352.2 354.5 355.7 355.6 356.9 359.4 361.5 363.8 360.4 354.3 357.2 363.9 372.4 382.6 399.1 402.9 400.6 395.5 390.5 388.9 390.2 391.1 391.9 391.5 390.4 390.2 391.3 391.6 391.0 388.6 386.2 389.1 403.8 430.2 441.8 449.5 448.1 443.2 438.3 432.9 430.7 434.6 442.0 447.6 446.0 440.3 434.7 431.1 435.7 442.3 445.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 423.8 402.1 398.5 397.6 393.4 390.5 391.4 402.5 427.6 442.5 442.5 435.1 430.1 430.8 439.9 447.4 442.2 426.1 412.3 399.8 391.8 389.5 388.5 387.8 386.4 384.7 384.5 387.9 391.2 391.8 392.9 393.8 392.0 392.0 395.4 398.1 398.2 396.3 393.1 391.1 388.9 386.6 383.1 381.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 357.8 367.6 370.9 365.4 359.3 355.7 358.1 372.7 396.8 404.0 398.4 392.6 389.2 388.8 383.6 362.3 341.1 325.1 326.3 327.9 331.3 333.3 326.0 319.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 356.4 353.5 349.0 346.5 338.4 328.4 323.7 326.2 334.2 338.4 331.6 309.8 280.6 256.4 252.8 256.1 258.3 256.9 257.6 261.1 259.7 258.7 258.6 260.2 262.2 262.4 262.9 264.2 263.4 260.2 256.4 253.2 238.7 223.3 231.6 257.7 258.1 258.4 258.8 258.0 256.8 255.8 254.0 255.1 258.1 261.9 263.7 262.6 256.7 253.1 250.2 246.7 258.7 294.4 327.1 342.6 346.0 344.1 341.2 342.3 345.2 350.1 364.7 382.5 396.1 396.1 389.2 381.8 381.1 387.2 397.0 399.3 390.7 374.3 360.1 350.6 346.6 347.4 350.7 354.4 354.3 351.7 349.8 348.4 346.9 347.0 348.4 349.5 351.0 352.3 353.6 353.3 350.5 348.3 345.5 344.3 344.4 347.0 350.6 352.0 351.0 350.8 350.1 347.6 345.7 347.0 350.3 351.7 350.7 348.5 346.9 347.7 349.0 349.1 348.5 346.8 346.3 348.4 349.0 349.2 351.1 349.6 348.3 350.5 351.1 348.0 347.6 349.1 351.3 356.0 361.3 360.6 354.0 341.0 316.2 302.9 302.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0"
+  }
+]
@@ -1,211 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Benchmark legacy and automatic MP4 response encoding on synthetic input."""
-
-from __future__ import annotations
-
-import argparse
-import hashlib
-import json
-import statistics
-import time
-from collections.abc import Callable
-from dataclasses import dataclass
-from pathlib import Path
-
-import numpy as np
-
-from vllm_omni.entrypoints.openai.video_api_utils import (
-    _encode_video_bytes,
-    _encode_video_bytes_legacy,
-)
-
-
-@dataclass(frozen=True)
-class BenchmarkVariant:
-    label: str
-    encoder: Callable[..., bytes]
-
-
-def _positive_int(value: str) -> int:
-    parsed = int(value)
-    if parsed < 1:
-        raise argparse.ArgumentTypeError("value must be at least 1")
-    return parsed
-
-
-def _build_inputs(
-    *,
-    frames: int,
-    height: int,
-    width: int,
-    fps: int,
-    audio_sample_rate: int,
-    seed: int,
-) -> tuple[np.ndarray, np.ndarray]:
-    rng = np.random.default_rng(seed)
-    planar = rng.random((3, frames, height, width), dtype=np.float32)
-    video = planar.transpose(1, 2, 3, 0)
-    if not all(video[..., channel].flags.c_contiguous for channel in range(3)):
-        raise RuntimeError("synthetic video does not expose contiguous channel planes")
-
-    sample_count = max(1, round(frames / fps * audio_sample_rate))
-    timeline = np.arange(sample_count, dtype=np.float32) / audio_sample_rate
-    left = np.sin(2 * np.pi * 220 * timeline).astype(np.float32, copy=False)
-    right = np.sin(2 * np.pi * 330 * timeline).astype(np.float32, copy=False)
-    return video, np.stack((left, right))
-
-
-def _measure(
-    variant: BenchmarkVariant,
-    *,
-    video: np.ndarray,
-    audio: np.ndarray,
-    fps: int,
-    audio_sample_rate: int,
-) -> tuple[bytes, dict[str, object]]:
-    cpu_start = time.process_time_ns()
-    wall_start = time.perf_counter_ns()
-    output = variant.encoder(
-        video,
-        fps=fps,
-        audio=audio,
-        audio_sample_rate=audio_sample_rate,
-        video_codec_options={"preset": "ultrafast", "threads": "0"},
-    )
-    wall_ms = (time.perf_counter_ns() - wall_start) / 1_000_000
-    process_cpu_ms = (time.process_time_ns() - cpu_start) / 1_000_000
-    return output, {
-        "label": variant.label,
-        "wall_ms": wall_ms,
-        "process_cpu_ms": process_cpu_ms,
-        "output_bytes": len(output),
-        "output_sha256": hashlib.sha256(output).hexdigest(),
-    }
-
-
-def _summarize(records: list[dict[str, object]], label: str) -> dict[str, object]:
-    selected = [record for record in records if record["label"] == label]
-    wall_values = [float(record["wall_ms"]) for record in selected]
-    cpu_values = [float(record["process_cpu_ms"]) for record in selected]
-    return {
-        "runs": len(selected),
-        "wall_ms": {
-            "median": statistics.median(wall_values),
-            "min": min(wall_values),
-            "max": max(wall_values),
-        },
-        "process_cpu_ms": {
-            "median": statistics.median(cpu_values),
-            "min": min(cpu_values),
-            "max": max(cpu_values),
-        },
-    }
-
-
-def _parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("--frames", type=_positive_int, default=12)
-    parser.add_argument("--height", type=_positive_int, default=96)
-    parser.add_argument("--width", type=_positive_int, default=160)
-    parser.add_argument("--fps", type=_positive_int, default=24)
-    parser.add_argument("--audio-sample-rate", type=_positive_int, default=32000)
-    parser.add_argument("--warmup", type=int, default=1)
-    parser.add_argument("--rounds", type=_positive_int, default=3)
-    parser.add_argument("--seed", type=int, default=20260817)
-    parser.add_argument("--output-json", type=Path)
-    args = parser.parse_args()
-    if args.warmup < 0:
-        parser.error("--warmup must be non-negative")
-    return args
-
-
-def main() -> None:
-    args = _parse_args()
-    video, audio = _build_inputs(
-        frames=args.frames,
-        height=args.height,
-        width=args.width,
-        fps=args.fps,
-        audio_sample_rate=args.audio_sample_rate,
-        seed=args.seed,
-    )
-    baseline = BenchmarkVariant("legacy", _encode_video_bytes_legacy)
-    candidate = BenchmarkVariant("automatic", _encode_video_bytes)
-
-    for _ in range(args.warmup):
-        for variant in (baseline, candidate):
-            _measure(
-                variant,
-                video=video,
-                audio=audio,
-                fps=args.fps,
-                audio_sample_rate=args.audio_sample_rate,
-            )
-
-    records: list[dict[str, object]] = []
-    for round_index in range(args.rounds):
-        order = (baseline, candidate) if round_index % 2 == 0 else (candidate, baseline)
-        for variant in order:
-            _, record = _measure(
-                variant,
-                video=video,
-                audio=audio,
-                fps=args.fps,
-                audio_sample_rate=args.audio_sample_rate,
-            )
-            record["round"] = round_index + 1
-            records.append(record)
-
-    output_hashes = {str(record["output_sha256"]) for record in records}
-    if len(output_hashes) != 1:
-        raise RuntimeError(f"benchmark variants produced different outputs: {sorted(output_hashes)}")
-
-    paired_savings_ms = []
-    paired_savings_percent = []
-    for round_index in range(args.rounds):
-        pair = [record for record in records if record["round"] == round_index + 1]
-        baseline_ms = float(next(record["wall_ms"] for record in pair if record["label"] == baseline.label))
-        candidate_ms = float(next(record["wall_ms"] for record in pair if record["label"] == candidate.label))
-        paired_savings_ms.append(baseline_ms - candidate_ms)
-        paired_savings_percent.append((baseline_ms - candidate_ms) / baseline_ms * 100)
-
-    result = {
-        "config": {
-            "frames": args.frames,
-            "height": args.height,
-            "width": args.width,
-            "fps": args.fps,
-            "audio_sample_rate": args.audio_sample_rate,
-            "warmup": args.warmup,
-            "rounds": args.rounds,
-            "seed": args.seed,
-            "video_shape": list(video.shape),
-            "video_strides": list(video.strides),
-            "codec_options": {"preset": "ultrafast", "threads": "0"},
-        },
-        "records": records,
-        "summary": {
-            baseline.label: _summarize(records, baseline.label),
-            candidate.label: _summarize(records, candidate.label),
-            "paired_wall_saving_ms": {
-                "median": statistics.median(paired_savings_ms),
-                "min": min(paired_savings_ms),
-                "max": max(paired_savings_ms),
-            },
-            "paired_wall_saving_percent": {
-                "median": statistics.median(paired_savings_percent),
-                "min": min(paired_savings_percent),
-                "max": max(paired_savings_percent),
-            },
-            "output_sha256": output_hashes.pop(),
-        },
-    }
-    output_json = json.dumps(result, indent=2, sort_keys=True)
-    if args.output_json is not None:
-        args.output_json.write_text(output_json + "\n", encoding="utf-8")
-    print(output_json)
-
-
-if __name__ == "__main__":
-    main()
@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 from __future__ import annotations

 import sys
@@ -62,23 +61,16 @@ def test_bootstrap_steps_loaded_from_file() -> None:


 def test_docs_only_allows_main_scheduled_nightly_weekly_only() -> None:
-    """skip_all: no PR labels; main + NIGHTLY=1 / WEEKLY=1 / NON_CRITICAL=1 still gates scheduled CI."""
+    """skip_all: no PR labels; main + NIGHTLY=1 / WEEKLY=1 still gates scheduled CI."""
    rendered = _render(["docs/foo.md"])
    assert "key: image-build" in rendered
    assert "key: upload-nightly-pipeline" in rendered
    assert "key: upload-weekly-pipeline" in rendered
-    # Scheduled main+WEEKLY=1 uploads L2/L3 with --e2e; NIGHTLY still gates L4 only.
-    doc = yaml.safe_load(rendered)
-    by_key = {step["key"]: step for step in doc["steps"]}
-    assert "NIGHTLY" not in by_key["upload-ready-pipeline"]["if"]
-    assert 'build.branch == "main"' in by_key["upload-ready-pipeline"]["if"]
-    assert 'build.env("WEEKLY") == "1"' in by_key["upload-ready-pipeline"]["if"]
-    assert "NIGHTLY" not in by_key["upload-merge-pipeline"]["if"]
-    assert 'build.branch == "main"' in by_key["upload-merge-pipeline"]["if"]
-    assert 'build.env("WEEKLY") == "1"' in by_key["upload-merge-pipeline"]["if"]
-    assert 'build.env("NIGHTLY") == "1"' in by_key["upload-nightly-pipeline"]["if"]
+    # Scheduled nightly also uploads L2/L3 with --e2e
+    assert "key: upload-ready-pipeline" in rendered
+    assert "key: upload-merge-pipeline" in rendered
+    assert 'build.env("NIGHTLY") == "1"' in rendered
    assert 'build.env("WEEKLY") == "1"' in rendered
-    assert 'build.env("NON_CRITICAL") == "1"' in rendered
    assert "nightly-test" not in rendered
    assert "weekly-test" not in rendered
    assert "merge-test" not in rendered
@@ -86,19 +78,6 @@ def test_docs_only_allows_main_scheduled_nightly_weekly_only() -> None:
    assert "if: false" not in rendered


-def test_npu_docs_only_does_not_upload_ready_on_nightly() -> None:
-    """NPU skip_all: scheduled NIGHTLY still uploads L4, not L2 ready."""
-    rendered = _render_bootstrap_pipeline(
-        BOOTSTRAP_STEPS_TEMPLATE,
-        decision=resolve_ci_decision(["docs/foo.md"]),
-        path=Path(".buildkite/npu/bootstrap-upload-steps.yml"),
-    )
-    doc = yaml.safe_load(rendered)
-    by_key = {step["key"]: step for step in doc["steps"]}
-    assert "upload-ready-pipeline" not in by_key
-    assert 'build.env("NIGHTLY") == "1"' in by_key["upload-nightly-pipeline"]["if"]
-
-
 def test_yaml_gated_l45_only_does_not_unconditionally_build_image() -> None:
    rendered = _render([".buildkite/cuda/test-nightly.yml"])
    assert "if: true" not in rendered
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Unit tests for StageConfigFactory and related classes.
 """
@@ -1267,6 +1267,8 @@ stages:
    @pytest.mark.parametrize(
        ("deploy_name", "pipeline_name", "stage_count", "final_output_type", "declared_pipeline"),
        [
+            ("mammoth_moda2.yaml", "mammoth_moda2", 2, "image", "mammoth_moda2"),
+            ("mammoth_moda2_ar.yaml", "mammoth_moda2_ar", 1, "text", "mammoth_moda2_ar"),
            ("omnivoice.yaml", "omnivoice", 1, "audio", "omnivoice"),
            ("mimo_audio.yaml", "mimo_audio", 2, "audio", None),
            ("step_audio_2.yaml", "step_audio_2", 2, "audio", None),
@@ -1304,7 +1306,6 @@ stages:
            ({"model_type": "step_audio_2"}, None, "step_audio_2"),
            (None, {"_class_name": "HunyuanVideo15Pipeline"}, "hunyuan_video_15"),
            (None, {"_class_name": "WanPipeline"}, "wan2_2_ti2v"),
-            (None, {"_class_name": "WanDMDPipeline"}, "wan2_2_ti2v"),
        ],
    )
    def test_migrated_models_are_discovered_without_explicit_deploy(
@@ -1660,7 +1661,6 @@ stages:
            ]
        }
        original_foo = fake_config["stages"][0]["foo"]
-        assert isinstance(original_foo, dict)
        original_b = original_foo["b"]

        with patch("vllm_omni.config.stage_config.resolve_deploy_yaml", return_value=fake_config):
@@ -2128,12 +2128,7 @@ class TestBaseConfigInheritance:
            sampling = deploy.stages[1].default_sampling_params
            assert sampling is not None, f"{filename} stage 1 lost default_sampling_params"
            assert sampling["temperature"] == 0.8, filename
-            # Upstream utils.TTSSamplingParams. min_tokens is not checked here:
-            # the duplex overlay drops it to 0 because the model budgets each
-            # generate_chunk itself.
-            assert sampling["top_k"] == 25, filename
-            assert sampling["top_p"] == 0.85, filename
-            assert sampling["repetition_penalty"] == 1.05, filename
+            assert sampling["top_k"] == 100, filename

    def test_ci_inherits_from_main(self):
        ci_path = Path(get_deploy_config_path("ci/qwen3_omni_moe.yaml"))
@@ -2225,28 +2220,6 @@ class TestPlatformOverrides:
        assert rocm.stages[0].enforce_eager is None
        assert rocm.stages[1].enforce_eager is True

-    def test_qwen3_omni_cuda_uses_thinker_rotary_custom_op(self):
-        deploy_path = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
-        pipeline = resolve_pipeline_config(
-            "qwen3_omni_moe",
-            Q3_OMNI_ALL_STAGES_HF_CONFIG,
-        )
-        assert isinstance(pipeline, PipelineConfig)
-
-        cuda = _apply_platform_overrides(load_deploy_config(deploy_path), platform="cuda")
-        cuda_stages = merge_pipeline_deploy(pipeline, cuda)
-        expected = {"custom_ops": ["+rotary_embedding"]}
-        assert cuda_stages[0].yaml_engine_args["compilation_config"] == expected
-        assert all("compilation_config" not in stage.yaml_engine_args for stage in cuda_stages[1:])
-
-        for platform in ("cpu", "musa", "npu", "rocm", "xpu"):
-            deploy = _apply_platform_overrides(
-                load_deploy_config(deploy_path),
-                platform=platform,
-            )
-            config = deploy.stages[0].compilation_config or {}
-            assert "+rotary_embedding" not in config.get("custom_ops", [])
-
    def test_minicpmo_4_5_cuda_caps_talker_kv_cache(self):
        pipeline = resolve_pipeline_config("minicpmo_4_5")
        assert isinstance(pipeline, PipelineConfig)
@@ -2753,23 +2726,22 @@ class TestSentinelDefaultPrecedence:
        )

    def test_ming_flash_omni_topology(self):
-        """Guard ming_flash_omni's SIP cleanup: stage 0 has no
-        ``custom_process_next_stage_input_func``, and stage 1 does not set
-        ``requires_full_payload_input=True``. Stage 1 uses only
-        ``thinker2talker_token_only`` (sync_process_input_func).
+        """Guard ming_flash_omni's SIP cleanup: stage 0 has no full-payload
+        producer hook (arch is not in ``_FULL_PAYLOAD_INPUT_STAGES``), and
+        stage 1 uses only ``thinker2talker_token_only`` (sync_process_input_func).
        The dead ``thinker2talker`` (custom_process_input_func) was removed
        because ming_flash_omni has no async_chunk support and both functions
        called ``_build_talker_inputs`` identically.
-        Merge under either async_chunk mode must not re-introduce either
-        full-payload transport gate."""
+        Merge under either async_chunk mode must not re-introduce a
+        stage-0 full-payload hook."""
        pipeline = resolve_pipeline_config("ming_flash_omni")
        assert isinstance(pipeline, PipelineConfig)

        stage0, stage1 = pipeline.stages
        assert stage0.custom_process_next_stage_input_func is None, (
-            "ming_flash_omni stage 0 must not declare a full-payload producer hook."
+            "ming_flash_omni stage 0 must not declare a full-payload producer "
+            "(connector path is not active for this arch)."
        )
-        assert stage1.requires_full_payload_input is False
        assert stage1.custom_process_input_func is None
        assert stage1.sync_process_input_func is not None
        assert stage1.sync_process_input_func.endswith("thinker2talker_token_only")
@@ -2856,14 +2828,13 @@ class TestObjectStorageConfigResolution:
        monkeypatch.setattr(config_factory_module, "ObjectStorageModel", FakeObjectStorageModel)

        class Recorder:
-            def __init__(self) -> None:
-                self.pulls = pulls
-
            def set_files(self, files: list[tuple[str, str]]) -> None:
                materialized_files.clear()
                materialized_files.extend(files)

-        return Recorder()
+        recorder = Recorder()
+        recorder.pulls = pulls
+        return recorder

    def test_passthrough_for_non_uri(self, fake_object_storage):
        assert _materialize_object_storage_configs("org/model") == "org/model"
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Tests for the additive structured Omni config."""

 from __future__ import annotations
@@ -244,36 +244,6 @@ def test_from_pipeline_config_rejects_unowned_deploy_engine_extras(engine_extras
        VllmOmniConfig.from_pipeline_config(pipeline, user_deploy_config=deploy)


-@pytest.mark.parametrize(
-    ("engine_extras", "cli_overrides"),
-    [
-        ({"requires_full_payload_input": False}, {}),
-        ({}, {"stage_1_requires_full_payload_input": False}),
-    ],
-    ids=["deploy", "stage-cli"],
-)
-def test_from_pipeline_config_rejects_full_payload_input_capability_overrides(
-    engine_extras,
-    cli_overrides,
-):
-    pipeline = _resolve_pipeline_or_skip("qwen3_tts")
-    deploy = DeployConfig(
-        stages=[
-            StageDeployConfig(
-                stage_id=1,
-                engine_extras=engine_extras,
-            )
-        ]
-    )
-
-    with pytest.raises(ValueError, match=r"no structured config owner: requires_full_payload_input"):
-        VllmOmniConfig.from_pipeline_config(
-            pipeline,
-            user_deploy_config=deploy,
-            cli_overrides=cli_overrides,
-        )
-
-
@pytest.mark.parametrize(
    ("cli_overrides", "stage_id"),
    [
@@ -473,12 +443,10 @@ def test_from_pipeline_config_dispatches_async_chunk_processors_without_mutating
    async_config = _from_pipeline_key("qwen3_tts")
    assert async_config.stage_by_id(0).custom_process_next_stage_input_func.endswith("talker2code2wav_async_chunk")
    assert async_config.stage_by_id(1).custom_process_input_func is None
-    assert async_config.stage_by_id(1).model_config.requires_full_payload_input is True

    sync_config = _from_pipeline_key("qwen3_tts", cli_overrides={"async_chunk": False})
    assert sync_config.stage_by_id(0).custom_process_next_stage_input_func.endswith("talker2code2wav_full_payload")
    assert sync_config.stage_by_id(1).custom_process_input_func.endswith("talker2code2wav_token_only")
-    assert sync_config.stage_by_id(1).model_config.requires_full_payload_input is True

    assert pipeline.get_stage(0).custom_process_next_stage_input_func.endswith("talker2code2wav_full_payload")
    assert pipeline.get_stage(1).custom_process_input_func is None
@@ -599,7 +567,6 @@ def test_sub_config_fields_match_structured_scopes():
        # legacy engine-args path.
        "model_subdir",
        "tokenizer_subdir",
-        "requires_full_payload_input",
    }
    vllm_load_fields = {f.name for f in fields(VllmLoadConfig)}
    assert issubclass(OmniStageLoadConfig, VllmLoadConfig)
@@ -1123,10 +1090,7 @@ def test_diffusion_config_preserves_existing_coercion_hooks():


 def test_diffusion_config_from_kwargs_reuses_legacy_normalization(monkeypatch):
-    from vllm_omni.platforms import current_omni_platform
-
    monkeypatch.setenv("DIFFUSION_CACHE_BACKEND", "TEA_CACHE")
-    monkeypatch.setattr(current_omni_platform, "is_cuda", lambda: True)

    cfg = omni_config_module._DiffusionConfigProjection.from_kwargs(
        diffusion_attention_backend="flash_attn",
@@ -1136,7 +1100,6 @@ def test_diffusion_config_from_kwargs_reuses_legacy_normalization(monkeypatch):
        kv_cache_skip_layers=[2],
        static_lora_scale=0.25,
        diffusion_kv_mode="paged_scheduler",
-        diffusion_kv_max_rows_per_request=2,
        diffusers_load_kwargs=None,
        diffusers_call_kwargs=None,
    )
@@ -1153,10 +1116,7 @@ def test_diffusion_config_from_kwargs_reuses_legacy_normalization(monkeypatch):
    assert cfg.diffusers_call_kwargs == {}


-def test_from_pipeline_config_normalizes_diffusion_config_aliases_from_engine_args(tmp_path, monkeypatch):
-    from vllm_omni.platforms import current_omni_platform
-
-    monkeypatch.setattr(current_omni_platform, "is_cuda", lambda: True)
+def test_from_pipeline_config_normalizes_diffusion_config_aliases_from_engine_args(tmp_path):
    deploy_path = tmp_path / "dreamzero_diffusion_aliases.yaml"
    deploy_path.write_text(
        "\n".join(
@@ -1168,7 +1128,6 @@ def test_from_pipeline_config_normalizes_diffusion_config_aliases_from_engine_ar
                "    diffusion_attention_backend: flash_attn",
                "    fa_deterministic: true",
                "    diffusion_kv_mode: paged_scheduler",
-                "    diffusion_kv_max_rows_per_request: 2",
            ]
        )
    )
@@ -1182,36 +1141,6 @@ def test_from_pipeline_config_normalizes_diffusion_config_aliases_from_engine_ar
    assert stage.diffusion_config.diffusion_attention_config.default.backend == "flash_attn"
    assert stage.diffusion_config.fa_deterministic is True
    assert stage.diffusion_config.diffusion_kv_mode is DiffusionKVCacheMode.PAGED_SCHEDULER
-    assert stage.diffusion_config.diffusion_kv_max_rows_per_request == 2
-
-
-def test_from_pipeline_config_forwards_fastvideo_vsa_topk(tmp_path, monkeypatch):
-    from vllm_omni.platforms import current_omni_platform
-
-    monkeypatch.setattr(current_omni_platform, "is_cuda", lambda: True)
-    deploy_path = tmp_path / "dreamzero_fastvideo_vsa.yaml"
-    deploy_path.write_text(
-        "\n".join(
-            [
-                "pipeline: dreamzero",
-                "async_chunk: false",
-                "stages:",
-                "  - stage_id: 0",
-                "    diffusion_attention_backend: FASTVIDEO_VSA",
-                "    fastvideo_vsa_topk: 96",
-            ]
-        )
-    )
-
-    stage = _from_pipeline_key(
-        "dreamzero",
-        deploy_config_path=str(deploy_path),
-    ).stage_by_id(0)
-
-    attention_config = stage.diffusion_config.diffusion_attention_config
-    assert attention_config.default is not None
-    assert attention_config.default.backend == "FASTVIDEO_VSA"
-    assert attention_config.default.fastvideo_vsa_topk == 96


 def test_from_pipeline_config_rejects_reserved_diffusion_kv_mode(tmp_path):
@@ -1254,7 +1183,6 @@ def test_diffusion_config_field_classification_covers_current_fields():
        "prompt_embed_cache_size",
        "diffusion_kv_cache_dtype",
        "diffusion_kv_mode",
-        "diffusion_kv_max_rows_per_request",
    } <= omni_config_module._DIFFUSION_ONLY_CONFIG_FIELDS
    assert {
        "revision",
@@ -1,10 +1,7 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
-
 """
 Root pytest entrypoint for the vLLM-Omni test suite.

- `tests/conftest.py` stays thin: plugin registration only.
+- `tests/conftest.py` stays thin: plugin registration + compatibility re-exports.
 - Importable utilities live under `tests/helpers/`.
 - Helper unit tests live under `tests/helpers/tests/`.
 - Fixtures live under `tests/helpers/fixtures/` and are loaded via `pytest_plugins`.
@@ -22,11 +19,12 @@ bootstrap_vllm_layer_custom_op_modules()

 pytest_plugins = (
    "tests.helpers.fixtures.config",
-    "tests.helpers.fixtures.clean",
+    "tests.helpers.fixtures.env",
    "tests.helpers.fixtures.log",
    "tests.helpers.fixtures.media",
    "tests.helpers.fixtures.run_args",
    "tests.helpers.fixtures.runtime",
+    "tests.helpers.fixtures.speaker_cache",
    "tests.helpers.fixtures.xdist",
 )

@@ -34,3 +32,46 @@ pytest_plugins = (
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
    # Marker for Buildkite log folding before pytest summary lines.
    terminalreporter.write_line("--- Running Summary")
+
+
+# Backward-compatible lazy re-exports.
+# (Many tests still import from `tests.conftest`; migrate these imports to `tests.helpers.*` over time.)
+# Keep these lazy so conftest import does not trigger heavy helper dependencies.
+_ASSERTION_EXPORT_NAMES = (
+    "assert_audio_speech_response",
+    "assert_diffusion_response",
+    "assert_http_error",
+    "assert_image_diffusion_response",
+    "assert_image_valid",
+    "assert_omni_response",
+    "assert_video_diffusion_response",
+    "assert_video_valid",
+)
+_MEDIA_EXPORT_NAMES = (
+    "convert_audio_bytes_to_text",
+    "convert_audio_file_to_text",
+    "cosine_similarity_text",
+    "decode_b64_image",
+    "generate_synthetic_audio",
+    "generate_synthetic_image",
+    "generate_synthetic_video",
+)
+_STAGE_CONFIG_EXPORT_NAMES = ("modify_stage_config",)
+_RUNTIME_EXPORT_NAMES = (
+    "DiffusionResponse",
+    "HttpResponse",
+    "OmniResponse",
+    "OmniRunner",
+    "OmniRunnerHandler",
+    "OmniServer",
+    "OmniServerParams",
+    "OmniServerStageCli",
+    "OpenAIClientHandler",
+    "dummy_messages_from_mix_data",
+)
+_LAZY_EXPORT_MODULES = {
+    **{name: "tests.helpers.assertions" for name in _ASSERTION_EXPORT_NAMES},
+    **{name: "tests.helpers.media" for name in _MEDIA_EXPORT_NAMES},
+    **{name: "tests.helpers.stage_config" for name in _STAGE_CONFIG_EXPORT_NAMES},
+    **{name: "tests.helpers.runtime" for name in _RUNTIME_EXPORT_NAMES},
+}
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 """Unit tests for Omni AR streaming-session async placeholder handling."""

@@ -33,9 +33,6 @@ def _make_scheduler(*, stage_id: int = 0) -> OmniARScheduler:
    sched.log_stats = False
    sched.chunk_transfer_adapter = None
    sched.skipped_waiting = set()
-    sched._free_request_blocks = MagicMock()
-    sched.encoder_cache_manager = MagicMock()
-    sched._inflight_prefills = set()
    return sched


@@ -64,9 +61,6 @@ def _run_resumable_segment_stop(
    session: Request,
    *,
    session_finished: bool = False,
-    handle_stopped=None,
-    chunk_transfer_adapter=None,
-    inter_stage_output=None,
 ):
    sched = MagicMock()
    sched.requests = {session.request_id: session}
@@ -78,14 +72,11 @@ def _run_resumable_segment_stop(
        return [42], True

    sched._update_request_with_output.side_effect = stop_request
+    sched._handle_stopped_request.return_value = session_finished
    # vLLM 0.26 returns (kv_xfer_params, ec_xfer_params); an unconfigured
    # MagicMock iterates empty and fails to unpack at the call site.
    sched._free_request.return_value = (None, None)
-    if handle_stopped is None:
-        sched._handle_stopped_request.return_value = session_finished
-    else:
-        sched._handle_stopped_request.side_effect = handle_stopped
-    sched.chunk_transfer_adapter = chunk_transfer_adapter
+    sched.chunk_transfer_adapter = None
    sched.running = [session]
    sched.waiting_for_transfer_free = set()
    sched.transfer_triggered_requests = set()
@@ -111,7 +102,6 @@ def _run_resumable_segment_stop(
    model_runner_output.cudagraph_stats = None
    model_runner_output.req_id_to_index = {session.request_id: 0}
    model_runner_output.routed_experts = None
-    model_runner_output.inter_stage_outputs = [inter_stage_output] if inter_stage_output is not None else None

    return OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)

@@ -167,41 +157,6 @@ def test_update_from_output_settles_in_flight_tokens() -> None:
    assert session.num_in_flight_tokens == 0


-def test_resumable_segment_boundary_keeps_pre_transition_send_watermark() -> None:
-    """The old segment's transfer must not observe the next segment's reset.
-
-    A queued streaming update can replace the same mutable Request while the
-    stop output is still being handled. The connector needs the confirmed
-    token count from before that replacement to avoid dropping the boundary.
-    """
-    session = _make_request()
-    session.status = RequestStatus.RUNNING
-    session.resumable = True
-    session.num_computed_tokens = 26
-    adapter = MagicMock()
-    adapter._confirmed_num_computed_tokens.return_value = 26
-    inter_stage_output = {"codes": {"audio": [7]}}
-
-    def replace_with_next_segment(request: Request) -> bool:
-        request.num_computed_tokens = 0
-        request.status = RequestStatus.WAITING
-        return False
-
-    _run_resumable_segment_stop(
-        session,
-        handle_stopped=replace_with_next_segment,
-        chunk_transfer_adapter=adapter,
-        inter_stage_output=inter_stage_output,
-    )
-
-    adapter.save_async.assert_called_once_with(
-        inter_stage_output,
-        session,
-        True,
-        confirmed_num_computed_tokens=26,
-    )
-
-
 def test_running_decode_step_without_inter_stage_payload_does_not_raise() -> None:
    """A decode step that neither stops nor carries an inter-stage payload.

@@ -321,33 +276,29 @@ def test_explicit_streaming_payload_replaces_placeholder_prompt() -> None:
        "meta": {"turn_eos_token_id": 99},
    }
    assert session.status == RequestStatus.WAITING
-    sched._free_request_blocks.assert_called_once_with(session)
-    sched.encoder_cache_manager.free.assert_called_once_with(session)


-def test_explicit_model_intermediate_prompt_replacement_releases_cache_and_watermark() -> None:
+def test_model_intermediate_streaming_payload_replaces_computed_prompt() -> None:
    sched = _make_scheduler(stage_id=1)
-    session = _make_request()
    sched.chunk_transfer_adapter = SimpleNamespace(
        receives_chunks=False,
        segment_finished_requests=set(),
-        requests_num_chunks_sent={session.external_req_id: 59},
    )
+    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
-    session.num_in_flight_tokens = 2
    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": {
-            "next_stage_prompt_len": 10,
            "replace_streaming_prompt": True,
+            "next_stage_prompt_len": 10,
        },
    }

@@ -357,42 +308,6 @@ def test_explicit_model_intermediate_prompt_replacement_releases_cache_and_water
    assert list(session._all_token_ids) == [0] * 10
    assert session.num_prompt_tokens == 10
    assert session.num_computed_tokens == 0
-    assert session.num_stale_output_tokens == 2
    assert session.additional_information is None
    assert session.model_intermediate_buffer == update.model_intermediate_buffer
    assert session.status == RequestStatus.WAITING
-    assert sched.chunk_transfer_adapter.requests_num_chunks_sent == {}
-    sched._free_request_blocks.assert_called_once_with(session)
-    sched.encoder_cache_manager.free.assert_called_once_with(session)
-
-
-def test_ready_async_chunk_prompt_replacement_releases_stale_kv_once() -> None:
-    sched = _make_scheduler(stage_id=1)
-    session = _make_request()
-    session.external_req_id = "external-ar-streaming-test"
-    session.num_in_flight_tokens = 2
-    # _update_request_as_session() may have already fenced this frame before
-    # the connector marks the explicit replacement ready.
-    session.num_stale_output_tokens = 2
-    session.num_output_placeholders = 2
-    session.spec_token_ids = [-1, -1]
-    sched.requests = {session.request_id: session}
-    sched._inflight_prefills.add(session)
-    sched.chunk_transfer_adapter = SimpleNamespace(
-        replaced_streaming_prompt_ids={session.request_id},
-        requests_with_ready_chunks={session.request_id},
-        requests_num_chunks_sent={session.external_req_id: 4090},
-    )
-
-    sched._reset_ready_async_chunk_replacements()
-    sched._reset_ready_async_chunk_replacements()
-
-    sched._free_request_blocks.assert_called_once_with(session)
-    sched.encoder_cache_manager.free.assert_called_once_with(session)
-    assert session not in sched._inflight_prefills
-    assert session.num_stale_output_tokens == 2
-    assert session.num_output_placeholders == 0
-    assert session.spec_token_ids == []
-    assert sched.chunk_transfer_adapter.replaced_streaming_prompt_ids == set()
-    assert sched.chunk_transfer_adapter.requests_with_ready_chunks == {session.request_id}
-    assert sched.chunk_transfer_adapter.requests_num_chunks_sent == {}
@@ -4,8 +4,10 @@ from types import SimpleNamespace
 import pytest
 from vllm.v1.engine import FinishReason

-from vllm_omni.core.sched import omni_scheduler_mixin
-from vllm_omni.core.sched.omni_scheduler_mixin import OmniSchedulerMixin
+from vllm_omni.core.sched.omni_scheduler_mixin import (
+    DEFAULT_INPUT_WAIT_TIMEOUT_S,
+    OmniSchedulerMixin,
+)
 from vllm_omni.core.sched.output import OmniChunkRecvHandle

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -15,77 +17,6 @@ class _Scheduler(OmniSchedulerMixin):
    pass


-def test_async_chunk_adapter_initializes_for_stage_zero_sender_and_stage_one_receiver(monkeypatch):
-    created_adapters = []
-
-    def adapter_factory(config):
-        adapter = SimpleNamespace(vllm_config=config)
-        created_adapters.append(adapter)
-        return adapter
-
-    monkeypatch.setattr(omni_scheduler_mixin, "OmniChunkTransferAdapter", adapter_factory)
-
-    def init_scheduler(**model_config):
-        scheduler = _Scheduler()
-        scheduler.vllm_config = SimpleNamespace(model_config=SimpleNamespace(**model_config))
-        scheduler._init_omni_io_scheduling_state()
-        return scheduler
-
-    producer = init_scheduler(
-        stage_id=0,
-        async_chunk=True,
-        requires_full_payload_input=False,
-        custom_process_next_stage_input_func="test.pipeline.produce_async_chunk",
-    )
-    receiver = init_scheduler(
-        stage_id=1,
-        async_chunk=True,
-        requires_full_payload_input=True,
-        custom_process_next_stage_input_func=None,
-    )
-    terminal_stage = init_scheduler(
-        stage_id=2,
-        async_chunk=True,
-        requires_full_payload_input=False,
-        custom_process_next_stage_input_func=None,
-    )
-
-    assert producer.chunk_transfer_adapter is not None
-    assert receiver.chunk_transfer_adapter is not None
-    assert terminal_stage.chunk_transfer_adapter is not None
-    assert receiver.input_coordinator is None
-    assert len(created_adapters) == 3
-
-
-@pytest.mark.parametrize(
-    ("stage_id", "async_chunk", "required", "enabled"),
-    [
-        (0, False, True, False),
-        (1, True, True, False),
-        (1, False, False, False),
-        (1, False, True, True),
-    ],
-)
-def test_full_payload_coordinator_matches_legacy_gate(monkeypatch, stage_id, async_chunk, required, enabled):
-    monkeypatch.setattr(
-        omni_scheduler_mixin,
-        "OmniChunkTransferAdapter",
-        lambda _config: SimpleNamespace(),
-    )
-    scheduler = _Scheduler()
-    scheduler.vllm_config = SimpleNamespace(
-        model_config=SimpleNamespace(
-            stage_id=stage_id,
-            async_chunk=async_chunk,
-            requires_full_payload_input=required,
-        )
-    )
-
-    scheduler._init_omni_io_scheduling_state()
-
-    assert (scheduler.input_coordinator is not None) is enabled
-
-
 def test_schedule_lifecycle_helpers_process_and_restore_both_input_paths():
    calls = []
    scheduler = _Scheduler()
@@ -127,7 +58,7 @@ def test_schedule_lifecycle_helpers_process_and_restore_both_input_paths():
        ("process", scheduler.waiting, scheduler.running, scheduler.requests),
        # The chunk deadline runs after chunks are applied, so a chunk that
        # arrived this cycle resets the clock before it is measured (R1.1).
-        ("chunk-timeouts", omni_scheduler_mixin.DEFAULT_INPUT_WAIT_TIMEOUT_S),
+        ("chunk-timeouts", DEFAULT_INPUT_WAIT_TIMEOUT_S),
        ("failed-sends",),
        ("restore-chunks", scheduler.waiting, scheduler.running, scheduler.requests),
        ("restore-full", scheduler.waiting),
@@ -20,6 +20,7 @@ import torch
 import vllm_omni.core.sched.omni_scheduling_coordinator as coord_mod
 from vllm_omni.core.sched.omni_scheduling_coordinator import (
    OmniSchedulingCoordinator,
+    uses_full_payload_input_coordinator,
 )

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -95,6 +96,109 @@ class MockQueue:
 # ------------------------------------------------------------------ #


+class TestFullPayloadCoordinatorSelection(unittest.TestCase):
+    """Tests for the (model_arch, model_stage) whitelist gate.
+
+    The init_omni_connectors arch allowlist is keyed by ``model_arch`` and
+    is a superset of the stages registered here -- consumer-wait stages
+    must be registered explicitly in ``_FULL_PAYLOAD_INPUT_STAGES``, while
+    the init allowlist covers both producer- and consumer-side runners.
+    These tests pin which ``(arch, stage)`` pairs the gate fires for today.
+    """
+
+    # Expected whitelist (model_arch, model_stage).  Hardcoded to avoid the
+    # tautology of importing _FULL_PAYLOAD_INPUT_STAGES and asserting it
+    # against itself; any drift between this matrix and the whitelist will
+    # fail loudly here.
+    EXPECTED_FULL_PAYLOAD_INPUT_STAGES: frozenset[tuple[str, str]] = frozenset(
+        {
+            ("Qwen3OmniMoeForConditionalGeneration", "talker"),
+            ("Qwen3OmniMoeForConditionalGeneration", "code2wav"),
+            ("Qwen2_5OmniForConditionalGeneration", "talker"),
+            ("Qwen2_5OmniForConditionalGeneration", "code2wav"),
+            ("CovoAudioForConditionalGeneration", "code2wav"),
+            ("MiMoAudioModel", "code2wav"),
+            ("Qwen3TTSCode2Wav", "code2wav"),
+            ("MiniCPMO45Code2Wav", "code2wav"),
+            # nemotron_voicechat: sync-path RVQ-VAE decode waits for the
+            # talker's full [frames, 31] code-stack payload.
+            ("NemotronVoiceChatCode2Wav", "code2wav"),
+            ("CosyVoice3Model", "cosyvoice3_code2wav"),
+            # audex: sync-path TTS decoder (async_chunk: false deploys) and
+            # the TTA XCodec1 stage (always sync full-payload).
+            ("AudexCode2Wav", "audex_code2wav"),
+            ("AudexXCodec1", "audex_xcodec"),
+            ("IndexTTS2S2MelDecoder", "indextts2_s2mel_decoder"),
+            ("IndexTTS25S2MelDecoder", "indextts2_5_s2mel_decoder"),
+            ("DyninOmniForConditionalGeneration", "token2image"),
+            ("DyninOmniForConditionalGeneration", "token2audio"),
+        }
+    )
+
+    def test_whitelist_matches_expected_matrix(self):
+        """_FULL_PAYLOAD_INPUT_STAGES must equal the hardcoded expected matrix.
+
+        Catches both accidental additions (which would silently enable the
+        consumer-wait gate for a new arch) and accidental removals (which
+        would silently disable an enabled arch).
+        """
+        from vllm_omni.core.sched.omni_scheduling_coordinator import _FULL_PAYLOAD_INPUT_STAGES
+
+        self.assertEqual(
+            frozenset(_FULL_PAYLOAD_INPUT_STAGES),
+            self.EXPECTED_FULL_PAYLOAD_INPUT_STAGES,
+            msg="_FULL_PAYLOAD_INPUT_STAGES drifted from the expected matrix; "
+            "update EXPECTED_FULL_PAYLOAD_INPUT_STAGES if intentional.",
+        )
+
+    def test_all_whitelisted_arch_stage_pairs_fire_gate(self):
+        """Every (arch, stage) pair in the expected matrix must fire
+        the gate when stage_id > 0 and async_chunk=False.
+        """
+        for arch, stage in self.EXPECTED_FULL_PAYLOAD_INPUT_STAGES:
+            model_config = SimpleNamespace(
+                stage_id=1,
+                async_chunk=False,
+                model_arch=arch,
+                model_stage=stage,
+            )
+            self.assertTrue(
+                uses_full_payload_input_coordinator(model_config),
+                msg=f"expected gate to fire for {arch}/{stage}",
+            )
+
+    def test_other_arch_or_stage_or_mode_does_not_fire(self):
+        cases = [
+            SimpleNamespace(
+                stage_id=1, async_chunk=True, model_arch="Qwen3OmniMoeForConditionalGeneration", model_stage="talker"
+            ),
+            SimpleNamespace(
+                stage_id=0, async_chunk=False, model_arch="Qwen3OmniMoeForConditionalGeneration", model_stage="thinker"
+            ),
+            SimpleNamespace(
+                stage_id=1,
+                async_chunk=False,
+                model_arch="Qwen3OmniMoeForConditionalGeneration",
+                model_stage="some_future_stage",
+            ),
+            SimpleNamespace(
+                stage_id=1, async_chunk=False, model_arch="Qwen3TTSForConditionalGeneration", model_stage="code2wav"
+            ),
+            SimpleNamespace(
+                stage_id=1, async_chunk=False, model_arch="MingFlashOmniForConditionalGeneration", model_stage="talker"
+            ),
+            SimpleNamespace(stage_id=1, async_chunk=False, model_arch=None, model_stage="talker"),
+            SimpleNamespace(
+                stage_id=1, async_chunk=False, model_arch="Qwen3OmniMoeForConditionalGeneration", model_stage=None
+            ),
+        ]
+        for model_config in cases:
+            self.assertFalse(
+                uses_full_payload_input_coordinator(model_config),
+                msg=f"expected gate OFF for {model_config}",
+            )
+
+
 class TestCoordinatorUpdateRequestMetadata(unittest.TestCase):
    """Test update_request_metadata applies scheduling metadata to requests."""

@@ -105,7 +105,7 @@ def create_unique_server_pytest_params(
    configs: list[dict[str, Any]],
    stage_configs_dir: Path,
 ) -> list[Any]:
-    """Like :func:`_create_unique_server_params`, but wrap each row in ``pytest.param`` with JSON marks."""
+    """Like :func:`create_unique_server_params`, but wrap each row in ``pytest.param`` with JSON marks."""
    marks_by_name = _marks_by_test_name(configs)
    return [
        pytest.param(
@@ -113,7 +113,7 @@ def create_unique_server_pytest_params(
            marks=marks_by_name.get(row[0], []),
            id=row[0],
        )
-        for row in _create_unique_server_params(configs, stage_configs_dir)
+        for row in create_unique_server_params(configs, stage_configs_dir)
    ]


@@ -150,7 +150,7 @@ def create_paired_omni_benchmark_pytest_params(
    """Paired params for ``run_benchmark.py`` (omni/tts)."""
    mapping = create_test_parameter_mapping(configs)
    marks_by_name = _marks_by_test_name(configs)
-    server_entries = [(row, row[0]) for row in _create_unique_server_params(configs, stage_configs_dir)]
+    server_entries = [(row, row[0]) for row in create_unique_server_params(configs, stage_configs_dir)]
    params_by_test_name = {
        test_name: get_benchmark_params_for_server(test_name, mapping) for _, test_name in server_entries
    }
@@ -232,7 +232,7 @@ def _build_serve_args(serve_args: Any) -> list[str]:
    return args


-def _create_unique_server_params(
+def create_unique_server_params(
    configs: list[dict[str, Any]],
    stage_configs_dir: Path,
 ) -> list[tuple[str, str, str | None, str | None, tuple[str, ...], bool]]:
@@ -311,7 +311,7 @@ def create_reliability_omni_server_params(
    configs: list[dict[str, Any]], stage_configs_dir: Path
 ) -> list[OmniServerParams]:
    adjusted_configs = configs_with_platform_stage_configs(configs)
-    unique_params = _create_unique_server_params(adjusted_configs, stage_configs_dir)
+    unique_params = create_unique_server_params(adjusted_configs, stage_configs_dir)
    server_args_by_name = extract_server_args_by_test_name(adjusted_configs)
    return [
        OmniServerParams(
@@ -10,7 +10,7 @@
          "num_cards": 3
        }
      },
-      "slow",
+      "full_model",
      "omni"
    ],
    "server_params": {