Merge remote-tracking branch 'origin/main' into codex-api-server-user-guide

Signed-off-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> # Conflicts: # docs/user_guide/examples/online_serving/text_to_audio.md # examples/online_serving/stable_audio/README.md
[Core] Validate PipelineConfig in __post_init__ (#6829 )
2026-09-01 11:23:08 +08:00 · 2026-09-01 11:17:57 +08:00 · 2026-09-01 01:02:54 +00:00 · 2026-09-01 08:41:23 +08:00 · 2026-08-31 15:27:13 +00:00 · 2026-08-31 15:18:29 +00:00
421 changed files with 31549 additions and 3057 deletions
@@ -1,4 +1,7 @@
 #!/bin/bash
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 # vllm-omni customized version
 # Based on: https://github.com/vllm-project/ci-infra/blob/main/buildkite/bootstrap-amd.sh
 # Last synced: 2025-12-15
@@ -6,6 +9,9 @@
 set -euo pipefail
 # The bootstrap runs from the repository root; ShellCheck does not follow
 # sourced files unless invoked with -x.
 # shellcheck disable=SC1091
 source .buildkite/common/scripts/resolve_skip_ci.sh
 if [[ -z "${RUN_ALL:-}" ]]; then
@@ -63,6 +69,8 @@ upload_pipeline() {
    # Install minijinja
    ls .buildkite || buildkite-agent annotate --style error 'Please merge upstream main branch for buildkite CI'
    curl -sSfL https://github.com/mitsuhiko/minijinja/releases/download/2.3.1/minijinja-cli-installer.sh | sh
    # Installed by the minijinja bootstrap above and only present on the CI agent.
    # shellcheck disable=SC1091
    source /var/lib/buildkite-agent/.cargo/env
    if [[ $BUILDKITE_PIPELINE_SLUG == "fastcheck" ]]; then
@@ -90,22 +98,82 @@ upload_pipeline() {
    cd .buildkite/amd
    # Select test definition file: merge suite for main, ready suite for PRs.
-    # For debugging, DEBUG_TEST_YAML can override the selection — accepts
+    # For debugging, DEBUG_TEST_YAML accepts a comma-separated list containing
-    # "merge" or "ready" (case-insensitive).
+    # "merge" and/or "ready" (case-insensitive). Multiple suites are combined
    # before rendering so they share one amd-build step.
    if [[ -n "${DEBUG_TEST_YAML:-}" ]]; then
-        case "${DEBUG_TEST_YAML,,}" in
+        declare -a DEBUG_TEST_SPECS=()
-            merge)
+        declare -A SEEN_DEBUG_TESTS=()
-                TEST_YAML="test-amd-merge.yml"
+        IFS=',' read -ra REQUESTED_DEBUG_TESTS <<< "${DEBUG_TEST_YAML,,}"
-                ;;
+
-            ready)
+        for requested_test in "${REQUESTED_DEBUG_TESTS[@]}"; do
-                TEST_YAML="test-amd-ready.yml"
+            # Permit readable values such as "ready, merge," and ignore the
-                ;;
+            # empty item produced by a trailing comma.
-            *)
+            requested_test="${requested_test//[[:space:]]/}"
-                echo "ERROR: DEBUG_TEST_YAML must be 'merge' or 'ready', got '$DEBUG_TEST_YAML'" >&2
+            [[ -z "$requested_test" ]] && continue
            if [[ -n "${SEEN_DEBUG_TESTS[$requested_test]:-}" ]]; then
                echo "ERROR: duplicate DEBUG_TEST_YAML suite '$requested_test'" >&2
                exit 1
-                ;;
+            fi
-        esac
+            SEEN_DEBUG_TESTS[$requested_test]=1
-        echo "DEBUG_TEST_YAML override: using $TEST_YAML"
+
            case "$requested_test" in
                ready)
                    DEBUG_TEST_SPECS+=("READY_TESTS:test-amd-ready.yml")
                    ;;
                merge)
                    DEBUG_TEST_SPECS+=("MERGE_TESTS:test-amd-merge.yml")
                    ;;
                *)
                    echo "ERROR: DEBUG_TEST_YAML entries must be 'merge' or 'ready', got '$requested_test'" >&2
                    exit 1
                    ;;
            esac
        done
        if [[ ${#DEBUG_TEST_SPECS[@]} -eq 0 ]]; then
            echo "ERROR: DEBUG_TEST_YAML did not contain a test suite" >&2
            exit 1
        elif [[ ${#DEBUG_TEST_SPECS[@]} -eq 1 ]]; then
            TEST_YAML="${DEBUG_TEST_SPECS[0]#*:}"
        else
            TEST_YAML=$(mktemp "${TMPDIR:-/tmp}/amd-debug-tests.XXXXXX.yml")
            python - "$TEST_YAML" "${DEBUG_TEST_SPECS[@]}" <<'PY'
 import sys
 import yaml
 output_path, *suite_specs = sys.argv[1:]
 combined = {"env": {}, "steps": []}
 for suite_spec in suite_specs:
    group_name, input_path = suite_spec.split(":", 1)
    with open(input_path, encoding="utf-8") as test_file:
        suite = yaml.safe_load(test_file)
    for name, value in (suite.get("env") or {}).items():
        previous = combined["env"].get(name, value)
        if previous != value:
            raise ValueError(
                f"Conflicting environment value for {name}: {previous!r} != {value!r}"
            )
        combined["env"][name] = value
    suite_steps = []
    for entry in suite.get("steps") or []:
        if "group" in entry:
            suite_steps.extend(entry.get("steps") or [])
        else:
            suite_steps.append(entry)
    combined["steps"].append({"group": group_name, "steps": suite_steps})
 with open(output_path, "w", encoding="utf-8") as output_file:
    yaml.safe_dump(combined, output_file, sort_keys=False)
 PY
        fi
        echo "DEBUG_TEST_YAML override: using ${DEBUG_TEST_SPECS[*]}"
    elif [[ $BUILDKITE_BRANCH == "main" ]]; then
        TEST_YAML="test-amd-merge.yml"
    else
@@ -130,19 +198,20 @@ upload_pipeline() {
            > pipeline.yaml
    )
    cat pipeline.yaml
    if [[ "$TEST_YAML" == "${TMPDIR:-/tmp}/amd-debug-tests."*.yml ]]; then
        rm -f -- "$TEST_YAML"
    fi
    buildkite-agent artifact upload pipeline.yaml
    buildkite-agent pipeline upload pipeline.yaml
    exit 0
 }
 get_diff() {
-    $(git add .)
+    git diff --name-only --diff-filter=ACMDR "$(git merge-base origin/main HEAD)"
    echo $(git diff --name-only --diff-filter=ACMDR $(git merge-base origin/main HEAD))
 }
 get_diff_main() {
-    $(git add .)
+    git diff --name-only --diff-filter=ACMDR HEAD~1
    echo $(git diff --name-only --diff-filter=ACMDR HEAD~1)
 }
 file_diff=$(get_diff)
@@ -181,7 +250,7 @@ for file in $file_diff; do
    # First check if file matches any pattern
    matches_pattern=0
    for pattern in "${patterns[@]}"; do
-        if [[ $file == $pattern* ]] || [[ $file == $pattern ]]; then
+        if [[ $file == "$pattern"* ]] || [[ $file == "$pattern" ]]; then
            matches_pattern=1
            break
        fi
@@ -191,7 +260,7 @@ for file in $file_diff; do
    if [[ $matches_pattern -eq 1 ]]; then
        matches_ignore=0
        for ignore in "${ignore_patterns[@]}"; do
-            if [[ $file == $ignore* ]] || [[ $file == $ignore ]]; then
+            if [[ $file == "$ignore"* ]] || [[ $file == "$ignore" ]]; then
                matches_ignore=1
                break
            fi
@@ -21,7 +21,7 @@ steps:
    - label: "Simple · Diffusion Test · Shard %N/%t"
      agent_pool: mi300_1
-      parallelism: 2
+      parallelism: 4
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -71,26 +71,45 @@ steps:
      commands:
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - |
+        - pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
-          bash -c '
+        - >-
-            set +e
+          pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
-            pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
+          --deselect='tests/diffusion/offloader/test_diffusion_layerwise_offload.py::test_layerwise_offload_diffusion_model[stabilityai/stable-audio-open-1.0]'
            EXIT1=\$?
            pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
            EXIT2=\$?
            exit \$((EXIT1 | EXIT2))
          '
-    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
+    # TODO: Diagnose unstable Stable Audio layerwise-offload memory savings on ROCm.
    - label: "Diffusion Stable Audio Layerwise Offload Test (quarantined)"
      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: NonBlocking
      commands:
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
        - >-
          pytest -s -v
          'tests/diffusion/offloader/test_diffusion_layerwise_offload.py::test_layerwise_offload_diffusion_model[stabilityai/stable-audio-open-1.0]'
    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs) · Shard 1/2"
      agent_pool: mi300_4
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      timeout_in_minutes: 40
      grade: Blocking
      commands:
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
+        - pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
-        - timeout 20m pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
+
    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs) · Shard 2/2"
      agent_pool: mi300_4
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      timeout_in_minutes: 40
      grade: Blocking
      commands:
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
        - pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
    # merge-only tests
    - label: "Diffusion Tensor Parallelism Test"
@@ -171,21 +190,45 @@ steps:
      grade: Blocking
      commands:
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py -m 'advanced_model and cuda' --run-level 'advanced_model'
    - label: "Qwen3-TTS CustomVoice E2E Test · Shard %N/%t"
      agent_pool: mi300_1
-      parallelism: 2
+      parallelism: 3
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
      timeout_in_minutes: 80
      commands:
        - |
-          bash -c '
+          timeout --signal=TERM --kill-after=2m 80m bash -c '
            export VLLM_LOGGING_LEVEL=DEBUG
            export VLLM_WORKER_MULTIPROC_METHOD=spawn
            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model" --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+            TESTS=(
              "tests/e2e/online_serving/test_qwen3_tts_customvoice.py::test_default_cuda_graph_startup[async_chunk]"
              "tests/e2e/online_serving/test_qwen3_tts_customvoice.py::test_text_to_audio_002[async_chunk]"
              "tests/e2e/offline_inference/test_qwen3_tts_customvoice.py::test_text_to_audio_001[no_cuda_graph]"
            )
            TEST="$${TESTS[$$BUILDKITE_PARALLEL_JOB]}"
            pytest -s -v "$$TEST" -m "advanced_model" --run-level "advanced_model"
          '
    # TODO: Restore this case to the blocking sharded step once its long-running
    # Whisper validation on AMD CI is resolved.
    - label: "Qwen3-TTS CustomVoice async_chunk Test (quarantined)"
      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: NonBlocking
      timeout_in_minutes: 80
      commands:
        - |
          timeout --signal=TERM --kill-after=2m 75m bash -c '
            export VLLM_LOGGING_LEVEL=DEBUG
            export VLLM_WORKER_MULTIPROC_METHOD=spawn
            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
            pytest -s -v "tests/e2e/online_serving/test_qwen3_tts_customvoice.py::test_text_to_audio_001[async_chunk]" -m "advanced_model" --run-level "advanced_model"
          '
    - label: "Qwen3-TTS Base E2E Test"
@@ -191,10 +191,7 @@ steps:
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
-        - |
+        - pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"
          timeout 45m bash -c '
            pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"
          '
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
@@ -85,6 +85,9 @@ steps:
          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
          - vllm_omni/entrypoints/openai/serving_speech.py
          - vllm_omni/entrypoints/openai/tts_adapters/
          - vllm_omni/engine/stage_pool.py
        commands:
          - |
            timeout 40m bash -c "
@@ -104,6 +107,9 @@ steps:
          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
          - vllm_omni/entrypoints/openai/serving_speech.py
          - vllm_omni/entrypoints/openai/tts_adapters/
          - vllm_omni/engine/stage_pool.py
          - .buildkite/common/scripts/run_cov_split.sh
          - pyproject.toml
        # 40 -> 60: the coverage split runs the model twice.
@@ -122,6 +128,7 @@ steps:
      - label: "Omni · Qwen3-Omni Test"
        source_file_dependencies:
          - tests/e2e/offline_inference/test_qwen3_omni.py
          - tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py
          - tests/e2e/online_serving/test_qwen3_omni.py
          - tests/helpers/runtime.py
          - vllm_omni/entrypoints/openai/serving_chat.py
@@ -132,9 +139,10 @@ steps:
          - vllm_omni/model_executor/models/qwen3_tts/cuda_graph_decoder_wrapper.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
          - vllm_omni/deploy/qwen3_omni_moe.yaml
          - vllm_omni/deploy/qwen3_omni_moe_thinking.yaml
        timeout_in_minutes: 50
        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py -m 'advanced_model and cuda' --run-level 'advanced_model'
        mirror_hardwares: h100_2
      - label: "Omni · MiniCPM-o 4.5 Test"
@@ -166,6 +174,7 @@ steps:
          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
          - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_server_vad.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
          - vllm_omni/deploy/minicpmo_4_5.yaml
          - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
@@ -216,3 +225,15 @@ steps:
              pytest -s -v tests/e2e/offline_inference/test_wan22_t2v.py tests/e2e/online_serving/test_wan22_t2v.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
            "
        mirror_hardwares: h100_1
      - label: "Diffusion · HunyuanVideo-1.5 Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_hunyuan_video_15_expansion.py
          - vllm_omni/diffusion/models/hunyuan_video/
        commands:
          - |
            timeout 40m bash -c "
              export VLLM_IMAGE_FETCH_TIMEOUT=60
              pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
            "
        mirror_hardwares: h100_1
@@ -12,16 +12,22 @@ steps:
    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "omni-test"
    steps:
-      - label: ":full_moon: Omni · Function Test with H100"
+      - label: ":full_moon: Omni · Function Test with H100 · Single-GPU"
-        timeout_in_minutes: 120
+        timeout_in_minutes: 90
        commands:
-          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni and cards_1" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
        mirror_hardwares: h100_1
      - label: ":full_moon: Omni · Function Test with H100 · 2-GPU"
        timeout_in_minutes: 90
        commands:
          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni and cards_2" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
        mirror_hardwares: h100_2
      - label: ":full_moon: Omni · MiniCPM-o 4.5 Duplex Test"
        timeout_in_minutes: 50
        commands:
-          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni and cards_1" --run-level "full_model"
        mirror_hardwares: h100_1
      - label: ":full_moon: Omni · Doc Test with L4"
@@ -31,10 +37,10 @@ steps:
          - pytest -s -v tests/examples/ -m "full_model and omni and L4" --run-level "full_model"
        mirror_hardwares: l4_4
-      - label: ":full_moon: Omni · Doc Test with H100"
+      - label: ":full_moon: Omni · Doc Test with H100 · 2-GPU"
        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/examples/ -m "full_model and omni and H100" --run-level "full_model"
+          - pytest -s -v tests/examples/ -m "full_model and omni and H100 and cards_2" --run-level "full_model"
        mirror_hardwares: h100_2
      - label: ":full_moon: Omni · Accuracy Test"
@@ -136,15 +142,29 @@ steps:
          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
        mirror_hardwares: l4_1
-      - label: ":full_moon: TTS · Perf Test"
+      - label: ":full_moon: TTS · Perf Test · Single-GPU"
-        key: nightly-tts-performance
+        key: nightly-tts-performance-single
        timeout_in_minutes: 180
        commands:
          - export BENCHMARK_DIR=tests/dfx/perf/results
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
          - |
            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json -m "H100 and cards_1"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
            exit $$EXIT
        mirror_hardwares: h100_1
      - label: ":full_moon: TTS · Perf Test · 2-GPU"
        key: nightly-tts-performance-2gpu
        timeout_in_minutes: 180
        commands:
          - export BENCHMARK_DIR=tests/dfx/perf/results
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
          - |
            set +e
            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json -m "H100 and cards_2"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
            exit $$EXIT
@@ -179,7 +199,6 @@ steps:
      build.pull_request.labels includes "nightly-test" ||
      build.pull_request.labels includes "diffusion-x2iat-test"
    steps:
      # Use explicit file shards so single-card cases do not reserve 2+ GPUs.
      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Single-GPU"
        timeout_in_minutes: 120
@@ -187,24 +206,24 @@ steps:
          - >-
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
-            -m "full_model and diffusion and H100 and not distributed_cuda"
+            -m "full_model and diffusion and H100 and cards_1"
            --run-level "full_model"
        mirror_hardwares: h100_1
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Qwen-Image"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · 2-GPU"
        timeout_in_minutes: 120
        commands:
          - >-
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
+            -m "full_model and diffusion and H100 and cards_2"
            --run-level "full_model"
        mirror_hardwares: h100_2
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4 · Multi-GPU"
        timeout_in_minutes: 120
        commands:
-          - pytest -sv tests/e2e/ -k "not test_wan and not hunyuan" -m "full_model and diffusion and L4" --run-level "full_model" --ignore=tests/e2e/accuracy
+          - pytest -sv tests/e2e/ -k "not test_wan and not hunyuan" -m "full_model and diffusion and L4 and not cards_1" --run-level "full_model" --ignore=tests/e2e/accuracy
        mirror_hardwares: l4_4
      - label: ":full_moon: Diffusion X2I(&A&T) · Doc Test"
@@ -235,8 +254,8 @@ steps:
          - 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"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image · Single-GPU"
-        key: nightly-diffusion-x2iat-performance-qwen-image
+        key: nightly-diffusion-x2iat-performance-qwen-image-single
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
@@ -244,7 +263,23 @@ steps:
          - export CACHE_DIT_VERSION=1.5.0
          - |
            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json -m "H100 and cards_1"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
            exit $$EXIT
        mirror_hardwares: h100_1
      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image · 4-GPU"
        key: nightly-diffusion-x2iat-performance-qwen-image-4gpu
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
          - export CACHE_DIT_VERSION=1.5.0
          - |
            set +e
            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json -m "H100 and cards_4"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
@@ -260,16 +295,34 @@ steps:
      build.pull_request.labels includes "diffusion-x2v-test"
    steps:
-      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test · Single-GPU"
        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda and H100 and cards_1" --run-level "full_model"
        mirror_hardwares: h100_1
      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test · 2-GPU"
        timeout_in_minutes: 90
        commands:
          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda and H100 and cards_2" --run-level "full_model"
        mirror_hardwares: h100_2
-      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test · Single-GPU"
        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda" --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda and H100 and cards_1" --run-level "full_model"
        mirror_hardwares: h100_1
      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test · 2-GPU"
        timeout_in_minutes: 90
        commands:
          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda and H100 and cards_2" --run-level "full_model"
        mirror_hardwares: h100_2
      - label: ":full_moon: Diffusion X2V · HunyuanVideo-1.5 Function Test"
        timeout_in_minutes: 90
        commands:
          - pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m "full_model and cuda" --run-level "full_model"
        mirror_hardwares: h100_2
      - label: ":full_moon: Diffusion X2V · Doc Test"
@@ -304,17 +357,49 @@ steps:
          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py -m full_model --run-level full_model
        mirror_hardwares: h100_2
-      - label: ":full_moon: Diffusion X2V · Perf Test"
+      - label: ":full_moon: Diffusion X2V · Perf Test · Single-GPU"
-        key: nightly-diffusion-x2v-performance
+        key: nightly-diffusion-x2v-performance-single
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
          - |
            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json -m "H100 and cards_1"
            EXIT=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
            exit $$EXIT
        mirror_hardwares: h100_1
      - label: ":full_moon: Diffusion X2V · Perf Test · 2-GPU"
        key: nightly-diffusion-x2v-performance-2gpu
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
          - |
            set +e
            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json -m "H100 and cards_2"
            EXIT1=$$?
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json -m "H100 and cards_2"
            EXIT2=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
            if [ $$EXIT1 -ne 0 ]; then exit $$EXIT1; fi
            exit $$EXIT2
        mirror_hardwares: h100_2
      - label: ":full_moon: Diffusion X2V · Perf Test · HunyuanVideo-1.5"
        key: nightly-diffusion-x2v-performance-hunyuanvideo15
        timeout_in_minutes: 180
        commands:
          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
          - |
            set +e
            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_hunyuanvideo15_t2v_vllm_omni.json
            EXIT1=$$?
            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_hunyuanvideo15_i2v_vllm_omni.json
            EXIT2=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
@@ -348,10 +433,16 @@ steps:
          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4' --run-level "full_model"
        mirror_hardwares: l4_4
-      - label: ":full_moon: Diffusion · Distributed Test with H100"
+      - label: ":full_moon: Diffusion · Distributed Test with H100 · 2-GPU"
        timeout_in_minutes: 60
        commands:
-          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100' --run-level "full_model"
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100 and cards_2' --run-level "full_model"
        mirror_hardwares: h100_2
      - label: ":full_moon: Diffusion · Distributed Test with H100 · 4-GPU"
        timeout_in_minutes: 60
        commands:
          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100 and cards_4' --run-level "full_model"
        mirror_hardwares: h100_4
      - label: ":full_moon: Diffusion · Offloader Test"
@@ -386,19 +477,25 @@ steps:
    depends_on:
      - nightly-omni-performance-no-async-chunk
      - nightly-omni-performance-async-chunk
-      - nightly-tts-performance
+      - nightly-tts-performance-single
-      - nightly-diffusion-x2iat-performance-qwen-image
+      - nightly-tts-performance-2gpu
-      - nightly-diffusion-x2v-performance
+      - nightly-diffusion-x2iat-performance-qwen-image-single
      - nightly-diffusion-x2iat-performance-qwen-image-4gpu
      - nightly-diffusion-x2v-performance-single
      - nightly-diffusion-x2v-performance-2gpu
    if: build.env("NIGHTLY") == "1" && build.env("EMAIL_DISTRIBUTION") == "1"
    commands:
      - pip install openpyxl
      - export DEFAULT_INPUT_DIR=tests/dfx/perf/results
      - export DEFAULT_OUTPUT_DIR=tests/dfx/perf/results
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-single
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-2gpu
      - 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-diffusion-x2iat-performance-qwen-image
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-single
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-4gpu
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance-single
      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance-2gpu
      - python tools/nightly/generate_nightly_perf_excel.py
      - python tools/nightly/generate_nightly_perf_html.py
      - python tools/nightly/send_nightly_email.py --report-file "tests/dfx/perf/results/*.xlsx, tests/dfx/perf/results/*.html"
@@ -205,6 +205,7 @@ steps:
          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
          - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_server_vad.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
          - vllm_omni/deploy/minicpmo_4_5.yaml
          - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
@@ -251,6 +252,9 @@ steps:
          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
          - vllm_omni/entrypoints/openai/serving_speech.py
          - vllm_omni/entrypoints/openai/tts_adapters/
          - vllm_omni/engine/stage_pool.py
        commands:
          - |
            timeout 20m bash -c "
@@ -259,25 +263,6 @@ steps:
            "
        mirror_hardwares: l4_1
      - label: "TTS · Qwen3-TTS Base Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_qwen3_tts_base.py
          - vllm_omni/model_executor/models/qwen3_tts/
          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
          - vllm_omni/model_executor/models/common/snake_activation.py
          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
        commands:
          - |
            timeout 20m bash -c "
              export VLLM_WORKER_MULTIPROC_METHOD=spawn
              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
              export VLLM_OMNI_USE_V2_RUNNER=1
              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
            "
        mirror_hardwares: l4_1
      - label: "Diffusion · Wan22 Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_wan22_t2v.py
@@ -337,3 +322,15 @@ steps:
              pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
            "
        mirror_hardwares: h100_1
      - label: "Diffusion · HunyuanVideo-1.5 Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_hunyuan_video_15_expansion.py
          - vllm_omni/diffusion/models/hunyuan_video/
        commands:
          - |
            timeout 40m bash -c "
              export VLLM_IMAGE_FETCH_TIMEOUT=60
              pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m 'core_model and diffusion' --run-level 'core_model'
            "
        mirror_hardwares: h100_1
@@ -44,10 +44,16 @@ steps:
          - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
        mirror_hardwares: h100_4
-      - label: "Reliability Test · Invalid parameters · H100"
+      - label: "Reliability Test · Invalid parameters · H100 · Single-GPU"
-        timeout_in_minutes: 120
+        timeout_in_minutes: 90
        commands:
-          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100"
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100 and cards_1"
        mirror_hardwares: h100_1
      - label: "Reliability Test · Invalid parameters · H100 · 2-GPU"
        timeout_in_minutes: 90
        commands:
          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100 and cards_2"
        mirror_hardwares: h100_2
      - label: "Reliability Test · Invalid parameters · L4"
@@ -105,27 +111,37 @@ steps:
    if: build.env("NON_CRITICAL") == "1"
    steps:
-      - label: "Omni · L4"
+      # Qwen2.5 expansion (CUDA 4-card / online 2-card) stays on this l4_4 job.
      - label: "Omni · L4 · Multi-GPU"
        timeout_in_minutes: 90
        commands:
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
+          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py -m "slow and L4 and omni and not cards_1" --run-level "core_model"
          - >-
            pytest -s -v tests/e2e/
            --ignore=tests/e2e/online_serving/test_qwen2_5_omni_expansion.py
            --ignore=tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py
-            -m "slow and L4 and omni"
+            -m "slow and L4 and omni and not cards_1"
            --run-level "full_model"
        mirror_hardwares: l4_4
-      - label: "Omni · H100"
+      - label: "Omni · H100 · Single-GPU"
        timeout_in_minutes: 90
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and H100 and omni"
+            tests/e2e/ -m "slow and H100 and omni and cards_1"
            --run-level "full_model"
-        mirror_hardwares: h100_2
+        mirror_hardwares: h100_1
      - label: "Omni · H100 · 4-GPU"
        timeout_in_minutes: 90
        commands:
          - >-
            pytest -sv
            tests/e2e/ -m "slow and H100 and omni and cards_4"
            --run-level "full_model"
        mirror_hardwares: h100_4
      - label: "TTS · L4"
        timeout_in_minutes: 120
@@ -169,12 +185,21 @@ steps:
            --run-level "full_model"
        mirror_hardwares: l4_1
-      - label: "Diffusion · L4"
+      - label: "Diffusion · L4 · Single-GPU"
        timeout_in_minutes: 60
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and diffusion and L4"
+            tests/e2e/ -m "slow and diffusion and L4 and cards_1"
            --run-level "full_model"
        mirror_hardwares: l4_1
      - label: "Diffusion · L4 · Multi-GPU"
        timeout_in_minutes: 60
        commands:
          - >-
            pytest -sv
            tests/e2e/ -m "slow and diffusion and L4 and not cards_1"
            --run-level "full_model"
        mirror_hardwares: l4_4
@@ -183,15 +208,24 @@ steps:
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and diffusion and H100 and not distributed_cuda"
+            tests/e2e/ -m "slow and diffusion and H100 and cards_1"
            --run-level "full_model"
        mirror_hardwares: h100_1
-      - label: "Diffusion · H100 · Multi-GPU"
+      - label: "Diffusion · H100 · 2-GPU"
        timeout_in_minutes: 90
        commands:
          - >-
            pytest -sv
-            tests/e2e/ -m "slow and diffusion and H100 and distributed_cuda"
+            tests/e2e/ -m "slow and diffusion and H100 and cards_2"
            --run-level "full_model"
        mirror_hardwares: h100_2
      - label: "Diffusion · H100 · 4-GPU"
        timeout_in_minutes: 90
        commands:
          - >-
            pytest -sv
            tests/e2e/ -m "slow and diffusion and H100 and cards_4"
            --run-level "full_model"
        mirror_hardwares: h100_4
@@ -10,7 +10,7 @@ steps:
          DOCKER_BUILDKIT: "1"
          # Buildkite will automatically replace this with the actual commit hash
          VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
-          VLLM_VERSION: "v0.27.0"
+          VLLM_VERSION: "v0.28.0"
        priority: 100
        timeout_in_minutes: 60
        soft_fail: false
@@ -5,7 +5,12 @@ set -ex
 omni_source_dir=$(git rev-parse --show-toplevel)
-base_image_name="xpu/vllm-omni-ci-base:${VLLM_VERSION:?VLLM_VERSION must be set}"
+: "${VLLM_VERSION:?VLLM_VERSION must be set}"
 # Official upstream XPU image. It is published from vllm's own
 # docker/Dockerfile.xpu, which is exactly what build_vllm_base() clones and
 # builds below, so pulling it is equivalent to building the fallback base.
 upstream_base_image="vllm/vllm-openai-xpu:${VLLM_VERSION}"
 local_base_image="xpu/vllm-omni-ci-base:${VLLM_VERSION}"
 image_name="xpu/vllm-omni-ci:${BUILDKITE_COMMIT:?BUILDKITE_COMMIT must be set}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(
    tr -dc A-Za-z0-9 </dev/urandom | head -c 10
@@ -49,9 +54,36 @@ build_vllm_base() (
        "${vllm_source_dir}"
 )
-if [ -z "$(docker images -q "${base_image_name}")" ]; then
+# Resolve the vLLM base image, in priority order:
-    build_vllm_base
+#   1. the published upstream image for VLLM_VERSION, pulled fresh (so moving
-fi
+#      tags like nightly are refreshed rather than served stale from disk);
 #   2. a local copy of that same image, when the registry is unreachable;
 #   3. VLLM_BASE, if that image is already on disk;
 #   4. otherwise build VLLM_BASE from the matching vLLM source tag (~30min).
 # VLLM_BASE names the fallback base and defaults to ${local_base_image}; set it
 # to reuse or produce a base under a different tag.
 resolve_vllm_base() {
    if docker pull "${upstream_base_image}"; then
        base_image_name="${upstream_base_image}"
        return
    fi
    # Registry unreachable, but a previous run may have left the image behind.
    if [ -n "$(docker images -q "${upstream_base_image}")" ]; then
        echo "WARNING: could not pull ${upstream_base_image}; using local copy" >&2
        base_image_name="${upstream_base_image}"
        return
    fi
    echo "WARNING: ${upstream_base_image} unavailable remotely and locally" >&2
    base_image_name="${VLLM_BASE:-${local_base_image}}"
    if [ -z "$(docker images -q "${base_image_name}")" ]; then
        echo "WARNING: building ${base_image_name} from vLLM ${VLLM_VERSION} source" >&2
        build_vllm_base
    fi
 }
 resolve_vllm_base
 # Try building the docker image
 docker_build "${image_name}" --build-arg "VLLM_BASE=${base_image_name}" --build-arg "VLLM_VERSION=${VLLM_VERSION}"
@@ -10,6 +10,18 @@ steps:
        -m 'core_model and npu and A3 and omni'
        --run-level 'core_model'
  - label: "Diffusion NPU layer tests"
    key: ready-diffusion-npu-layer-tests
    depends_on: upload-ready-pipeline
    mirror_hardwares: a3_npu_2
    commands:
      - >-
        pytest -s -v
        tests/diffusion/layers/test_activation.py
        tests/diffusion/layers/test_fused_qk_norm_rope_npu.py
        -m 'core_model and npu and A3'
        --run-level 'core_model'
  - group: ":card_index_dividers: Omni Model Test"
    key: ready-omni-group
    depends_on: upload-ready-pipeline
@@ -338,7 +338,7 @@ Classify the model's **CI priority** first:
 **L2 & L3 online — same file, dual marks on the baseline smoke:** The **first / simplest** case in `test_{slug}.py` (default deploy, minimal steps, single prompt) should carry **both** `@pytest.mark.core_model` and `@pytest.mark.advanced_model` on the **same** function so L2 (`test-ready.yml`) and L3 (`test-merge.yml`) share one smoke test. Heavier deploy variants or API paths in the same file use **`advanced_model` only**. When L3 moves to nightly, migrate those heavier cases into `test_{slug}_expansion.py` with `full_model` and remove the dedicated `test-merge.yml` job (see `test_longcat_image_expansion.py`, `test_qwen_image_expansion.py`).
-**L4 design (high priority):** Combine multiple supported features (Cache-DiT, TP, USP, CFG, HSDP, CPU offload, quantization) into **few parametrized** `OmniServerParams` rows so each feature appears in at least one case without exploding GPU jobs. Shard single-GPU vs multi-GPU cases across the nightly X2I/X2V function steps (1-GPU vs `distributed_cuda`).
+**L4 design (high priority):** Combine multiple supported features (Cache-DiT, TP, USP, CFG, HSDP, CPU offload, quantization) into **few parametrized** `OmniServerParams` rows so each feature appears in at least one case without exploding GPU jobs. Shard single-GPU vs multi-GPU cases across the nightly X2I/X2V function steps (`cards_1` vs `not cards_1`).
 **L4 design (medium / low):** One or two parametrized rows covering the best quality/perf trade-off; skip perf JSON unless the model is high priority.
@@ -34,7 +34,7 @@ but the changes are correct — stage the modified files and re-commit.
 | `check-forbidden-imports` | Stdlib `re`/`base64`, pickle, Hugging Face Hub API, or direct Triton/TileLang | `import regex as re` and `pybase64`; use `vllm.transformers_utils.repo_utils`. Do **not** add the file to `allowed_files` without review |
 | `check-torch-cuda-call` | New `torch.cuda.*` call site | Use `current_omni_platform` / `OmniPlatform`; do not grow `ALLOWED_FILES` without review |
 | `check-tts-adapter-migration` | New `self._tts_model_type` branch in `serving_speech.py` | Put per-model logic in `tts_adapters/`. Lower `MAX_MODEL_TYPE_BRANCHES` when removing branches; do not raise it without review |
-| `check-test-ci-coverage` | New `tests/**/test_*.py` missing level or hardware mark | Add `core_model`/`advanced_model`/… plus `cpu`/`cuda`/`hardware_test(` |
+| `check-mark` | New `tests/**/test_*.py` missing level/hardware mark, or direct `pytest.mark.H100` | Add `core_model`/`advanced_model`/… plus `cpu`/`cuda`/`hardware_test(`; SKU marks only via helpers |
 | `mypy-3.10` | Type error on changed `vllm_omni/` files | Fix the types; model trees are excluded. Extra versions: `pre-commit run --hook-stage manual mypy-3.12` |
 | `markdownlint-cli2` | Docs/README markdown lint | Accept auto-fix under `docs/` / `recipes/` / root README |
 | `check-buildkite` | Invalid `.buildkite/*.yml` | Fix the YAML; do not grow `SKIP_FILES` without review |
@@ -52,8 +52,8 @@ Always attach markers deliberately:
  - `tts` — speech synthesis / TTS-only models (`/v1/audio/speech`, voice clone, etc.)
  - `diffusion` — generative diffusion models (image / audio / text / video from noise)
 - **Cross-cutting area** (when relevant): `parallel`, `cache`, `example`, `benchmark`
- **Hardware**: `cpu`, `gpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `distributed_cuda`, …
+- **Hardware**: `cpu`, `gpu`, `cuda`, `rocm`, `npu`, …; SKU (`H100`, `L4`, …) and `cards_{n}` via `@hardware_test(...)` / `hardware_marks(...)` only — never `pytest.mark.H100` (`check-mark` rejects it)
- Optional: `slow`, distributed markers when multi-card is required
+- Optional: `slow`. Multi-card: `num_cards=N` on the helper (filter with `cards_N` / `not cards_1`)
 **Baseline smoke (L2 + L3):** The simplest e2e case per model — default deploy, minimal request — should usually carry **both** `@pytest.mark.core_model` and `@pytest.mark.advanced_model` on the **same** test function so `test-ready.yml` and `test-merge.yml` share one test. `send_*_request` picks validation depth from `--run-level`. References: `test_voxcpm2_tts.py::test_text_to_audio_001`, `test_qwen3_tts_customvoice.py::test_text_to_audio_001`. Heavier scenarios use **`advanced_model` only**; L4 expansion uses **`full_model`**.
@@ -274,7 +274,7 @@ def test_text_to_video_001(omni_server, online_client) -> None:
 **3. Cross-cutting rules**
- New `.py` files need the Omni SPDX header (`Copyright contributors to the vLLM-Omni project`, not `vLLM project`). `tests/` may keep stdlib `re`/`base64`; pickle is still banned unless the file is already on the pickle allowlist. Every collected `tests/**/test_*.py` needs a CI **level** mark and a **hardware** mark/helper (`check-test-ci-coverage` runs locally; GHA skips it). Do not grow `allowed_files` to land a test. Policy: [docs/contributing/README.md](../../../docs/contributing/README.md#linting).
+- New `.py` files need the Omni SPDX header (`Copyright contributors to the vLLM-Omni project`, not `vLLM project`). `tests/` may keep stdlib `re`/`base64`; pickle is still banned unless the file is already on the pickle allowlist. Every collected `tests/**/test_*.py` needs a CI **level** mark and a **hardware** platform mark/helper (`check-mark` runs locally; GHA skips it). Do not write `pytest.mark.H100` / `pytest.mark.L4`; use the helpers so `cards_{n}` is attached. Do not grow `allowed_files` to land a test. Policy: [docs/contributing/README.md](../../../docs/contributing/README.md#linting).
 - Reuse existing fixtures for the chosen scenario; do not mix “online client” assumptions into offline `OmniRunner` tests without a clear reason.
 - Avoid external network dependency in assertions unless the scenario is explicitly “online serving” or doc examples that require a model hub (then align with CI secrets/cache).
 - Keep **one test function = one intent** (one modality combo, one endpoint contract, or one acceleration combo).
@@ -79,7 +79,7 @@ E2E (L2+) should not rely on mocks unless documenting a rare exception; prefer r
   - `full_model` — L4 nightly (`test-nightly.yml`); some expansion tests still carry both `advanced_model` and `full_model` during migration
 2. **Model type** (pick one): `omni`, `tts`, or `diffusion`
 3. **Cross-cutting** when relevant: `parallel`, `cache`, `example`, `benchmark`
-4. **Hardware**: `cpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `distributed_cuda`, …
+4. **Hardware**: `cpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `cards_1`, …
 5. Multi-card: `@hardware_test(...)` in `tests/helpers/mark.py`
 ## Command Templates
@@ -244,7 +244,7 @@ pytest -s -v dfx/reliability/invalid_param_test/test_invalid_image_generation.py
 ```bash
 cd tests
-pytest -s -v -m "core_model and distributed_cuda and L4" --run-level=core_model
+pytest -s -v -m "core_model and L4 and not cards_1" --run-level=core_model
 ```
 ### Concrete e2e paths (common in-tree)
@@ -27,4 +27,4 @@ jobs:
        # Don't run historical-debt hooks against the whole tree in CI yet.
        # They still run locally on changed files.
        env:
-          SKIP: check-test-ci-coverage,markdownlint-cli2,shellcheck,check-spdx-header,mypy-3.10
+          SKIP: check-mark,markdownlint-cli2,shellcheck,check-spdx-header,mypy-3.10
@@ -105,8 +105,8 @@ repos:
      <<: *mypy_common
      stages: [manual]
-    - id: check-test-ci-coverage
+    - id: check-mark
-      name: Ensure test files have pytestmarks
+      name: Ensure test files have CI marks (no direct SKU)
      entry: python tools/pre_commit/check_test_marks.py
      language: python
      types: [python]
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0
+ARG VLLM_ASCEND_TAG=v0.28.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 # WORKDIR /vllm-workspace/vllm
@@ -30,6 +30,7 @@ RUN python3 -m pip install -v --no-deps \
      "stepaudio2-minicpmo>=0.1.1" \
      "step-audio2>=1.0.0"
 ENV VLLM_TARGET_DEVICE=npu
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 ENTRYPOINT []
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0-a3
+ARG VLLM_ASCEND_TAG=v0.28.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 # WORKDIR /vllm-workspace/vllm
@@ -30,6 +30,7 @@ RUN python3 -m pip install -v --no-deps \
      "stepaudio2-minicpmo>=0.1.1" \
      "step-audio2>=1.0.0"
 ENV VLLM_TARGET_DEVICE=npu
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 ENTRYPOINT []
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0
+ARG VLLM_ASCEND_TAG=v0.28.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -65,6 +65,7 @@ RUN python3 -m pip install -v --no-deps \
      --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
      --default-timeout=60
 ENV VLLM_TARGET_DEVICE=npu
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 ENTRYPOINT []
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.27.0-a3
+ARG VLLM_ASCEND_TAG=v0.28.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -66,6 +66,7 @@ RUN python3 -m pip install -v --no-deps \
      --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
      --default-timeout=60
 ENV VLLM_TARGET_DEVICE=npu
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 ENTRYPOINT []
@@ -1,140 +1,33 @@
-# Argument to configure vllm base image if pre-built
+# Target vLLM version: v0.28.0. Keep VLLM_VERSION aligned with the target vLLM
-ARG VLLM_BASE=vllm-base
+# release on every rebase.
-
+#
-FROM intel/deep-learning-essentials:2025.3.2-0-devel-ubuntu24.04 AS vllm-base
+# This file layers vLLM-Omni on top of the published upstream XPU image, which
-
+# already ships a self-consistent XPU runtime (torch-xpu, oneCCL/UCX+NIXL, the
-WORKDIR /workspace/
+# pinned UMD/compute-runtime stack, and vLLM itself) built from
-
+# https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu. Docker
-ARG PYTHON_VERSION=3.12
+# pulls that base automatically, so there is nothing to build here beyond the
-ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"
+# Omni layers. Mirrors how docker/Dockerfile.ci consumes vllm/vllm-openai.
-
+#
-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
+# If the tag does not exist yet (targeting an unreleased vLLM commit), build the
-    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list
+# base from upstream's own Dockerfile and point VLLM_BASE at the result:
-
+#
-RUN apt clean && apt-get update -y && \
+#   git clone --depth 1 --branch $VLLM_VERSION https://github.com/vllm-project/vllm /tmp/vllm
-    apt-get install -y --no-install-recommends --fix-missing \
+#   docker build -t vllm-openai-xpu:local --target vllm-openai \
-    curl \
+#       -f /tmp/vllm/docker/Dockerfile.xpu /tmp/vllm
-    ffmpeg \
+#   docker build -f docker/Dockerfile.xpu --build-arg VLLM_BASE=vllm-openai-xpu:local .
-    git \
+#
-    libsndfile1 \
+# .buildkite/intel/scripts/run-xpu-test.sh automates that pull-then-build fallback.
-    libsm6 \
+ARG VLLM_BASE_IMAGE=vllm/vllm-openai-xpu
-    libxext6 \
+ARG VLLM_VERSION=v0.28.0
-    libgl1 \
+ARG VLLM_BASE=${VLLM_BASE_IMAGE}:${VLLM_VERSION}
    lsb-release \
    libaio-dev \
    numactl \
    wget \
    vim \
    python3.12 \
    python3.12-dev \
    python3-pip
 # Add oneAPI repo, pin oneAPI to 2025.3, then install pinned packages in one layer.
 RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
    printf '%s\n' \
    'Package: intel-oneapi-* intel-deep-learning-essentials* intel-pti*' \
    'Pin: version 2025.3*' \
    'Pin-Priority: 1001' \
    > /etc/apt/preferences.d/oneapi-2025.3.pref && \
    apt-get update -y && \
    apt-get install -y --no-install-recommends \
    intel-oneapi-compiler-dpcpp-cpp-2025.3 \
    intel-oneapi-mkl-devel-2025.3 \
    intel-oneapi-dnnl-devel-2025.3 && \
    rm -rf /var/lib/apt/lists/*
 # Install UMD with fixed version
 RUN mkdir neo && \
    cd neo && \
    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-core-2_2.24.8+20344_amd64.deb && \
    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-opencl-2_2.24.8+20344_amd64.deb && \
    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-ocloc_25.48.36300.8-0_amd64.deb && \
    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-opencl-icd_25.48.36300.8-0_amd64.deb && \
    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libigdgmm12_22.8.2_amd64.deb && \
    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libze-intel-gpu1_25.48.36300.8-0_amd64.deb && \
    wget https://github.com/oneapi-src/level-zero/releases/download/v1.26.0/level-zero_1.26.0+u24.04_amd64.deb && \
    dpkg -i *.deb && \
    cd .. && \
    rm -rf neo
 ENV PATH="/root/.local/bin:$PATH"
 ENV VIRTUAL_ENV="/opt/venv"
 ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
 RUN curl -LsSf https://astral.sh/uv/install.sh | sh
 RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 # This oneccl contains the BMG support which is not the case for default version of oneapi 2025.3.
 ARG ONECCL_INSTALLER="intel-oneccl-2021.15.9.14_offline.sh"
 RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.9/${ONECCL_INSTALLER}" && \
    bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
    rm "${ONECCL_INSTALLER}" && \
    echo "source /opt/intel/oneapi/setvars.sh --force" >> /root/.bashrc && \
    echo "source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force" >> /root/.bashrc
 RUN rm -f /opt/intel/oneapi/ccl/latest && \
    ln -s /opt/intel/oneapi/ccl/2021.15 /opt/intel/oneapi/ccl/latest
 SHELL ["bash", "-c"]
 CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
 WORKDIR /workspace/
 ENV UV_HTTP_TIMEOUT=500
 # Configure package index for XPU
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE="copy"
 ARG VLLM_VERSION=v0.27.0
 RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
 WORKDIR /workspace/vllm
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --upgrade pip && \
    uv pip install -r requirements/xpu.txt && \
    uv pip install setuptools_scm &&\
    uv pip install grpcio-tools protobuf nanobind && \
    source /opt/intel/oneapi/setvars.sh --force && \
    source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force && \
    export CMAKE_PREFIX_PATH="$(python3 -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH}"
 ENV LD_LIBRARY_PATH="/opt/intel/oneapi/ccl/2021.15/lib:$LD_LIBRARY_PATH:/usr/local/lib/"
 ENV VLLM_TARGET_DEVICE=xpu
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip uninstall triton triton-xpu && \
    uv pip install triton-xpu==3.7.2 && \
    uv pip install --no-build-isolation --no-deps .
 CMD ["/bin/bash"]
 FROM vllm-base AS vllm-openai
 # install nixl from source code
 RUN uv pip install nixl
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install accelerate hf_transfer pytest pytest_asyncio lm_eval[api] modelscope
 # install development dependencies (for testing)
 RUN uv pip install -e tests/vllm_test_utils
 # ensure vllm is properly installed
 RUN python -c "import vllm, inspect; print(vllm.__file__)"
 RUN uv pip show vllm
 CMD ["/bin/bash"]
 ENTRYPOINT []
 FROM ${VLLM_BASE} AS vllm-omni
 # The base image ships ENTRYPOINT ["vllm", "serve"]; clear it so this stage stays
 # usable as a plain shell/test container regardless of which VLLM_BASE was used.
 ENTRYPOINT []
 CMD ["/bin/bash"]
 RUN apt-get update && \
    apt-get install -y espeak-ng ffmpeg git jq && \
    apt-get clean && \
@@ -5,14 +5,22 @@ nav:
    - getting_started/quickstart.md
    - getting_started/installation/*
  - Serving:
-    - OpenAI-Compatible API:
+    - API Server:
-      - Chat Completions API: serving/chat_completions_api.md
+      - Overview: serving/README.md
-      - Image Generation: serving/image_generation_api.md
+      - Core APIs:
-      - Image Edit: serving/image_edit_api.md
+        - Chat Completions: serving/chat_completions_api.md
-      - Text to Speech: serving/speech_api.md
+        - Text to Speech: serving/speech_api.md
-      - Videos: serving/videos_api.md
+        - Audio Generation: serving/audio_generate_api.md
-      - Audio Generation: serving/audio_generate_api.md
+        - Image Generation: serving/image_generation_api.md
-      - Streaming Video Input: serving/video_stream_api.md
+        - Image Edit: serving/image_edit_api.md
        - Videos: serving/videos_api.md
      - Streaming and Realtime:
        - Realtime Audio: serving/realtime_api.md
        - Full Duplex: serving/full_duplex_api.md
        - Streaming Video Input: serving/video_stream_api.md
        - Streaming Video Output: serving/streaming_video_output_api.md
        - OpenPI Robot Policy: serving/openpi_api.md
      - Standalone Experimental Servers: serving/standalone_servers.md
  - Examples:
    - examples/README.md
    - Offline Inference:
@@ -49,7 +49,7 @@ vLLM-Omni's `pre-commit` hooks will now run automatically every time you commit.
    `--all-files` run stays green while historical debt is cleaned up. The
    current `SKIP` list in
    [`.github/workflows/pre-commit.yml`](https://github.com/vllm-project/vllm-omni/blob/main/.github/workflows/pre-commit.yml)
-    is `check-test-ci-coverage`, `markdownlint-cli2`, `shellcheck`,
+    is `check-mark`, `markdownlint-cli2`, `shellcheck`,
    `check-spdx-header`, and `mypy-3.10`. Those hooks still run on **your
    commit** for changed files. A passing GitHub pre-commit check does not mean
    they passed locally. Hooks **not** on that list (forbidden imports,
@@ -66,7 +66,7 @@ already there). `check-pickle-imports` is gone: pickle is now one rule inside
 | `markdownlint-cli2` | Markdown in `docs/`, `recipes/`, `README.md`, `CONTRIBUTING.md` (not `.claude/` / `.cursor/` / `CLAUDE.md`) | skipped |
 | `mypy-3.10` | Type-check changed `vllm_omni/` files (model trees excluded). `tests/` uses `--follow-imports skip` | skipped |
 | `mypy-3.11` / `3.12` / `3.13` | Same checker, extra Python versions | not installed; `pre-commit run --hook-stage manual mypy-3.12` |
-| `check-test-ci-coverage` | Every `tests/**/test_*.py` has a CI level mark and a hardware mark/helper | skipped |
+| `check-mark` | Every `tests/**/test_*.py` has a CI level mark and a hardware platform mark or helper; no direct `pytest.mark.H100` / SKU | skipped |
 | `check-tts-adapter-migration` | `self._tts_model_type` branches in `serving_speech.py` must not increase | runs |
 | `shellcheck` | `*.sh` quoting / undefined vars | skipped |
 | `check-spdx-header` | Omni SPDX header on `.py` / `.pyi` / `.sh` / `.rs` / `.proto` | skipped |
@@ -144,11 +144,13 @@ pre-commit run --hook-stage manual mypy-3.12
 #### Test CI marks
-`check-test-ci-coverage` requires each collected test module under `tests/` to
+`check-mark` requires each collected test module under `tests/` to
 have at least one CI **level** mark (`core_model`, `advanced_model`,
-`full_model`, `local_model`, or `slow`) and a **hardware** mark (`cpu`, `cuda`,
+`full_model`, `local_model`, or `slow`) and a **hardware** platform mark
-`H100`, …) or helper (`hardware_test(` / `hardware_marks(`). GitHub Actions
+(`cpu`, `cuda`, …) or helper (`hardware_test(` / `hardware_marks(`). SKU marks
-skips this hook; local commit does not. See the
+(`H100`, `L4`, …) must come from those helpers so `cards_{n}` is attached;
 direct `pytest.mark.H100` is rejected. GitHub Actions skips this hook; local
 commit does not. See the
 [test writing guide](./ci/test_writing_guide.md).
 #### TTS adapter ratchet
@@ -402,7 +402,7 @@ loads the model once per mode.
 Before splitting, confirm each half actually collects at least one test under the job's forwarded filters — pytest exits with code 5 ("no tests collected") if a half is empty, which the script reports as a failure, red-ing a job that used to pass as one combined invocation.
 Also check the tests' `num_cards` against the job's `mirror_hardwares` preset.
-`cuda_marks` turns `num_cards > 1` into `skipif(device_count() < num_cards)`, so a
+`_cuda_marks` turns `num_cards > 1` into `skipif(device_count() < num_cards)`, so a
 test asking for more GPUs than the preset provides is silently skipped and its
 coverage artifact comes back empty. Splitting such a job produces a file that
 looks fine and measures nothing — "Diffusion · Z Image Test" is the current
@@ -81,8 +81,8 @@ Supported form:
 | Form | Example | Effect |
 | ---- | ------- | ------ |
-| Array (single platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}}, "full_model", "diffusion"]` | Hardware marks from `hardware_marks(...)`; `num_cards` > 1 adds `distributed_*` (+ CUDA `skipif_cuda` when applicable). String entries become extra pytest marks. |
+| Array (single platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}}, "full_model", "diffusion"]` | Hardware marks from `hardware_marks(...)`; `num_cards` adds `cards_{n}` (including `cards_1` when omitted). String entries become extra pytest marks. |
-| Array (multi-platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100", "rocm": "MI325", "npu": "A2"}, "num_cards": {"cuda": 2, "rocm": 2, "npu": 1}}}, "full_model", "omni"]` | Same as above, but declares **multiple platforms** in one case. Filter examples: `-m "full_model and H100 and cuda"`, `-m "full_model and MI325 and rocm"`. |
+| Array (multi-platform) | `"mark": [{"hardware_marks": {"res": {"cuda": "H100", "rocm": "MI325"}, "num_cards": 2}}, "full_model", "omni"]` | Same as above, but declares **multiple platforms** on one item. All platforms must share the same `num_cards` (int or identical dict values). Different counts need `@hardware_test` or one `hardware_marks()` per `pytest.param`. Filter examples: `-m "full_model and H100 and cards_2 and cuda"`, `-m "full_model and MI325 and cards_2 and rocm"`. |
 Recommended for L4 perf cases:
@@ -157,7 +157,7 @@ pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
 ```
 <!-- markdownlint-disable-next-line MD051 -->
-See also [Markers for Tests](#markers-for-tests) for registered hardware markers (`H100`, `L4`, `cuda`, `distributed_cuda`, …).
+See also [Markers for Tests](#markers-for-tests) for registered hardware markers (`H100`, `L4`, `cuda`, `cards_1`, …).
 #### `server_params` Configuration
@@ -88,7 +88,7 @@ Failed jobs: 1/2
    ```bash
    cd tests
    pytest -s -v test_xxxx.py --run-level=core_model
-    pytest -s -v -m "core_model and distributed_cuda and L4" --run-level=core_model
+    pytest -s -v -m "core_model and L4 and not cards_1" --run-level=core_model
    ```
 === "L3 level"
@@ -119,7 +119,7 @@ Failed jobs: 1/2
    ```bash
    pytest -s -v test_xxxx.py --run-level=advanced_model
-    pytest -s -v -m "advanced_model and distributed_cuda and L4" --run-level=advanced_model
+    pytest -s -v -m "advanced_model and L4 and not cards_1" --run-level=advanced_model
    ```
 === "L4 level"
@@ -153,12 +153,12 @@ Failed jobs: 1/2
    ```bash
    cd tests
    pytest -s -v test_xxxx.py --run-level=full_model
-    pytest -s -v -m "full_model and distributed_cuda and L4" --run-level=full_model
+    pytest -s -v -m "full_model and L4 and not cards_1" --run-level=full_model
    pytest -s -v -m "full_model and (omni or tts) and H100" --run-level=full_model
    ```
    If you only want to run specific test cases on a particular platform, you can use:
    ```bash
-    pytest -s -v -m "full_model and distributed_cuda and L4"  --run-level=full_model
+    pytest -s -v -m "full_model and L4 and not cards_1"  --run-level=full_model
    ```
    Note: ``run_benchmark.py`` and ``run_diffusion_benchmark.py`` accept an optional ``--test-config-file``. If omitted, each loads every ``*.json`` under ``tests/dfx/perf/tests/`` (omni/tts vs diffusion split by ``is_diffusion_perf_config``) and pytest ``-m`` filters by each case's JSON ``mark``:
    ```bash
@@ -48,17 +48,12 @@ Defined in `pyproject.toml`:
 | `MI325`            | Tests that require MI325 GPU (AMD/ROCm) *                                                               |
 | `A2`               | Tests that require A2 NPU *                                                                             |
 | `A3`               | Tests that require A3 NPU *                                                                             |
-| `distributed_cuda` | Tests that require multi cards on CUDA platform *                                                       |
+| `cards_{n}`        | Tests that require *n* accelerator cards (`cards_1` … `cards_8`); auto-added from `num_cards` *         |
 | `distributed_rocm` | Tests that require multi cards on ROCm platform  *                                                      |
 | `distributed_npu`  | Tests that require multi cards on NPU platform  *                                                       |
 | `skipif_cuda`      | Skip if the num of CUDA cards is less than the required *                                               |
 | `skipif_rocm`      | Skip if the num of ROCm cards is less than the required *                                               |
 | `skipif_npu`       | Skip if the num of NPU cards is less than the required *                                                |
 | `slow`             | Slow tests (may skip in quick CI)                                                                       |
 | `benchmark`        | Benchmark tests (decorator on runner test functions; perf JSON uses `full_model` + type marker instead) |
 | `local_model`      | Tests requiring local / non-HF-hub model weights                                                        |
-\* Means those markers are auto-added by `@hardware_test` (parametrization decorator) or `hardware_marks` (only returning the list of marks for flexibility).
+\* Means those markers are auto-added by `@hardware_test` (parametrization decorator) or `hardware_marks` (only returning the list of marks for flexibility). Do **not** write `pytest.mark.H100` / `pytest.mark.L4` (or other SKU names) by hand: `check-mark` rejects that so `cards_{n}` cannot be skipped.
 #### Example usage for markers
@@ -81,25 +76,23 @@ def test_video_to_audio()
 This decorator is intended to make hardware-aware, cross-platform test authoring easier and more robust for CI/CD environments. The `hardware_test` decorator in `vllm-omni/tests/helpers/mark.py` performs the following actions:
 1. **Applies platform and resource markers**  
-   Adds the appropriate pytest markers for each specified hardware platform (e.g., `cuda`, `rocm`, `xpu`, `npu`) and resource type (e.g., `L4`, `H100`, `MI325`, `B60`, `A2`, `A3`).
+   Adds the appropriate pytest markers for each specified hardware platform (e.g., `cuda`, `rocm`, `xpu`, `npu`) and resource type (e.g., `L4`, `H100`, `H200`, `B200`, `MI325`, `B60`, `A2`, `A3`).
   ```python
   @pytest.mark.cuda
   @pytest.mark.L4
   ```
-2. **Handles multi-card (distributed) scenarios**  
+2. **Handles card-count filtering**  
-   For tests requiring multiple cards, it automatically adds distributed markers such as `distributed_cuda`, `distributed_rocm`, or `distributed_npu`.
+   Adds a filterable marker `cards_{n}` (including `cards_1` when `num_cards` is omitted / 1). Use this with the SKU and platform, e.g. `-m "H100 and cards_4 and cuda"` vs `-m "H100 and cards_2 and cuda"`. Any multi-card CUDA test is `-m "cuda and not cards_1"`.
-   ```python
+
-   @pytest.mark.distributed_cuda(num_cards=num_cards)
+   When `res` lists more than one platform **and** `num_cards` differs by platform, `@hardware_test` collects **one variant per platform** so those `cards_{n}` marks do not stack. `-m "H100 and cards_4 and cuda"` then selects only the CUDA/H100 4-card variant, not a ROCm 2-card sibling. The same count on every platform stays a single item.
   ```
 3. **Supports flexible card requirements**  
   Accepts `num_cards` as either a single integer for all platforms or as a dictionary with per-platform values. If not specified, defaults to 1 card per platform.
 4. **Integrates resource validation**  
-   On CUDA, adds a skip marker (`skipif_cuda`) if the system does not have the required number of devices.
+   On CUDA and XPU, when `num_cards > 1`, attaches pytest's built-in `skipif` if the machine has fewer devices than requested. This is not a named `skipif_*` marker and is not used for `-m` filtering.
   Support for `skipif_rocm` and `skipif_npu` will be implemented later.
 5. **Works with pytest filtering**  
-   Allows tests to be filtered and selected at runtime using standard pytest marker expressions (e.g., `-m "distributed_cuda and L4"`).
+   Allows tests to be filtered and selected at runtime using standard pytest marker expressions (e.g., `-m "H100 and cards_2"`, `-m "L4 and not cards_1"`).
 ##### Example usage for decorator
@@ -117,13 +110,14 @@ This decorator is intended to make hardware-aware, cross-platform test authoring
        num_cards=2,
    )
    ```
- `res` must be a dict; supported resources: CUDA (L4/H100), ROCm (MI325), XPU (B60), MUSA (S5000), NPU (A2/A3)
+- `res` must be a dict. SKU values come from `pyproject.toml` markers tagged `[hardware-resource]` plus the platform key (`[cuda]`, `[rocm]`, `[xpu]`, `[npu]`, `[musa]`).
 - `num_cards` can be int (all platforms) or dict (per platform); defaults to 1 when missing
 - Distributed markers (`distributed_cuda`, `distributed_rocm`, `distributed_npu`) are auto-added for multi-card cases
 - Filtering examples:
-    - CUDA only: `pytest -m "distributed_cuda and L4"`
+    - CUDA H100 4-card: `pytest -m "H100 and cards_4 and cuda"`
-    - ROCm only: `pytest -m "distributed_rocm and MI325"`
+    - CUDA H100 2-card: `pytest -m "H100 and cards_2 and cuda"`
-    - NPU only: `pytest -m "distributed_npu"`
+    - CUDA multi-card (any count): `pytest -m "L4 and not cards_1"`
    - ROCm multi-card: `pytest -m "MI325 and not cards_1"`
    - NPU multi-card: `pytest -m "npu and not cards_1"`
 #### Function: `hardware_marks`
@@ -179,37 +173,30 @@ If you want to add support for a new platform (e.g., "tpu" for a new accelerator
   [tool.pytest.ini_options]
   markers = [
       # ... existing markers ...
-       "tpu: Tests that require TPU device",
+       "tpu: [hardware-platform] Tests that require TPU device",
-       "TPU_V3: Tests that require TPU v3 hardware",
+       # Add [gpu] next to [hardware-platform] if this platform should also get pytest.mark.gpu.
-       "distributed_tpu: Tests that require multiple TPU devices",
+       "TPU_V3: [hardware-resource] [tpu] Tests that require TPU v3 hardware",
       "cards_1: [hardware-cards] Tests that require 1 accelerator card",
       "cards_2: [hardware-cards] Tests that require 2 accelerator cards",
   ]
   ```
   `check-mark` treats `[hardware-platform]` names (including `cpu`/`gpu`) as allowed `pytest.mark.<name>` values. `hardware_test(res=...)` still rejects `cpu`/`gpu` as `res` keys.
 2. **Implement a marker construction function for your platform** in `vllm-omni/tests/helpers/mark.py`:
   ```python
   # In vllm-omni/tests/helpers/mark.py
-   def tpu_marks(*, res: str, num_cards: int):
+   def _tpu_marks(*, res: str, num_cards: int):
-       test_platform = pytest.mark.tpu
+       _require_sku("tpu", res)
-       if res == "TPU_V3":
+       return [pytest.mark.tpu, getattr(pytest.mark, res), _cards_mark(num_cards)]
           test_resource = pytest.mark.TPU_V3
       else:
           raise ValueError(
               f"Invalid TPU resource type: {res}. Supported: TPU_V3")
       if num_cards == 1:
           return [test_platform, test_resource]
       else:
           test_distributed = pytest.mark.distributed_tpu(num_cards=num_cards)
           # Optionally: add skipif_tpu when implemented
           return [test_platform, test_resource, test_distributed]
   ```
-3. **Update `hardware_marks` to recognize your new platform**:
+3. **Register the builder** in `_marks_for_platform`:
    In the relevant place (see the `hardware_marks` implementation), add:
    ```python
-    if platform == "tpu":
+    builders = {
-        marks = tpu_marks(res=resource, num_cards=cards)
+        # ... existing platforms ...
        "tpu": _tpu_marks,
    }
    ```
-    (`hardware_test` calls `hardware_marks` internally, so both will pick up the change.)
+    (`hardware_test` / `hardware_marks` both go through `_marks_for_platform`.)
 4. **(Recommended) Add a test using your new markers**:
   ```python
   @hardware_test(
@@ -223,11 +210,11 @@ If you want to add support for a new platform (e.g., "tpu" for a new accelerator
 **Summary**:  
 - Add pytest markers for your new platform/resources  
- Implement a marker function (`xxx_marks`)  
+- Implement a private marker function (`_xxx_marks`)
- Plug into `hardware_marks`  
+- Register it in `_marks_for_platform`
- You're done: tests using `@hardware_test` or `hardware_marks` with your platform now automatically get the correct markers, distribution, and isolation!
+- You're done: tests using `@hardware_test` or `hardware_marks` with your platform now automatically get the correct markers and `cards_{n}` filtering.
-See code in `vllm-omni/tests/helpers/mark.py` for existing examples (`cuda_marks`, `rocm_marks`, `npu_marks`).
+See code in `vllm-omni/tests/helpers/mark.py` for existing examples (`_cuda_marks`, `_rocm_marks`, `_npu_marks`).
 ## Test case style
@@ -438,7 +438,7 @@ selects the scheduler and worker family:
 Key topology fields belong to `StagePipelineConfig`:
 | Field | Description |
-|-------|-------------|
+| ------- | ------------- |
 | `model_stage` | Logical stage name (`ar_stage`, `decoder`, etc.). |
 | `execution_type` | Runtime family for this stage. |
 | `input_sources` | Upstream stage IDs that provide input. |
@@ -960,7 +960,7 @@ Common failures and fixes:
 | `check-forbidden-imports` | Stdlib `re`/`base64`, pickle, Hugging Face Hub API, or direct Triton/TileLang | `import regex as re` and `pybase64`; do not grow the allowlist without review |
 | `check-torch-cuda-call` | New `torch.cuda.*` call site | Use `current_omni_platform`; do not grow `ALLOWED_FILES` without review |
 | `check-tts-adapter-migration` | New `_tts_model_type` branch in `serving_speech.py` | Put logic in `tts_adapters/`; lower `MAX_MODEL_TYPE_BRANCHES` when removing branches |
-| `check-test-ci-coverage` | Test file missing level or hardware mark | Add `core_model`/… plus `cpu`/`cuda`/`hardware_test(` |
+| `check-mark` | Test file missing level/hardware mark, or direct `pytest.mark.H100` | Add `core_model`/… plus `cpu`/`cuda`/`hardware_test(`; SKU marks only via helpers |
 | `shellcheck` | Native `shellcheck` missing, or script warning | Install via apt/dnf/brew (or `shellcheck.exe` on Windows). See [Linting](../README.md#linting) |
 | `mypy-3.10` / `markdownlint-cli2` / `check-buildkite` | Types, docs markdown, or Buildkite YAML | See [Linting](../README.md#linting) |
@@ -26,6 +26,8 @@ Legend: ✅ supported, ⚠️ compatibility path or limited validation, ❌ unsu
 | **TP > 1** | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. |
 | **HSDP** | ❌ Rejected to avoid double-sharding parameters. | ⚠️ Limited end-to-end coverage. |
 | **Per-tensor online FP8 linears** | ✅ Ordinary loader finalizes weights and scales before DLO sharding. | ✅ Ordinary loader retains complete rank-local tensors. |
 | **Online INT8 linears** | ✅ Ordinary loader finalizes the int8 weight (contiguous or transposed view) and fp32 scale before DLO sharding. | ✅ Ordinary loader retains complete rank-local tensors. |
 | **Online MXFP8 linears** | ✅ Ordinary loader finalizes the fp8 weight (contiguous) and e8m0 block scale (contiguous or transposed view) before DLO sharding. | ✅ Ordinary loader retains complete rank-local tensors. |
 | **Other online quantization methods** | ❌ Rejected until runtime packing and scale layouts are validated. | ⚠️ Allowed through the ordinary loader; validation is method-specific. |
 | **Model-level or standard layerwise CPU offload** | ❌ Disabled because DLO takes priority. | ❌ Disabled because DLO takes priority. |
 | **Resident leading layers** | ❌ Rejected. | ✅ Requires eligible resident paths in the model's `OffloadPlan`. |
@@ -438,11 +440,15 @@ requirements.
 Direct checkpoint mmap can back either transfer path. It is currently limited
 to proven TP1, non-HSDP, non-online-quantized layouts. Other layouts use the
-ordinary loader. Per-tensor online FP8 linears can use DLO AllGather after the
+ordinary loader. Per-tensor online FP8, online INT8, and online MXFP8 linears
-ordinary loader finalizes their runtime weights and scales; DLO then shards and
+can use DLO AllGather after the ordinary loader finalizes their runtime
-reconstructs those tensors with their recorded layouts. Other online methods
+weights and scales; DLO then shards and reconstructs those tensors with their
-must use `--dlo-no-use-allgather` or disable online quantization until their
+recorded layouts (they all keep plain 1-byte dtypes over ordinary strided
-runtime layouts are validated.
+views: online INT8 has an int8 weight plus fp32 scale, online MXFP8 has an
 fp8 weight plus e8m0 block scale, each with a contiguous or transposed-view
 finalized tensor, all covered by the same physical-order packing as online
 FP8). Other online methods must use `--dlo-no-use-allgather` or disable
 online quantization until their runtime layouts are validated.
 The Host Weight Runtime representation, publication, and no-AllGather consumer
 contracts are merged; see
@@ -459,8 +465,8 @@ Current source-level validation includes:
 - HSDP + DLO without AllGather acceptance at configuration level;
 - loader preflight fallback for TP, HSDP, online quantization, unknown custom
  loaders, missing keys, and shape/dtype mismatches;
- ordinary-loader fallback for per-tensor online FP8 linears followed by DLO
+- ordinary-loader fallback for per-tensor online FP8, online INT8, and online
-  sharding of finalized weights and scales;
+  MXFP8 linears followed by DLO sharding of finalized weights and scales;
 - exact loader-to-backend plan transfer and ordinary-loader fallback;
 - ordered publication hashing, overlapped durability, unordered parallel
  checksum fallback, and node-local source-digest reuse/invalidation;
@@ -22,59 +22,60 @@ We use **Qwen-Image** as the reference implementation for decode parallel, and *
 **VAE Patch Parallelism** is an acceleration technique for both **encoding** and **decoding**. Instead of processing the entire tensor at once, the tensor is:
-+ Split into multiple spatial tiles
+- Split into multiple spatial tiles
-+ Distributed across multiple ranks
+- Distributed across multiple ranks
-+ Encoded/Decoded in parallel
+- Encoded/Decoded in parallel
-+ Merged to reconstruct the final output
+- Merged to reconstruct the final output
 This approach:
-+ Distributes computation across multiple devices
+- Distributes computation across multiple devices
-+ Reduces peak memory usage per device
+- Reduces peak memory usage per device
-+ Accelerates encoding/decoding latency
+- Accelerates encoding/decoding latency
 ### When to Use Encode vs Decode Parallel
-| Operation | Use Case | Example |
+| Operation           | Use Case                     | Example                           |
-|-----------|----------|---------|
+| ------------------- | ---------------------------- | --------------------------------- |
-| **Decode Parallel** | Text-to-Image, Text-to-Video | Latent → Image/Video |
+| **Decode Parallel** | Text-to-Image, Text-to-Video | Latent → Image/Video              |
-| **Encode Parallel** | Image-to-Video (I2V) | Image → Latent (for conditioning) |
+| **Encode Parallel** | Image-to-Video (I2V)         | Image → Latent (for conditioning) |
 ### Architecture
 We introduce **DistributedVaeExecutor** as the core component responsible for distributed VAE encoding/decoding.
 The executor is model-agnostic and accepts three function parameters:
-+ split – Partition the latent into tiles
+- split – Partition the latent into tiles
-+ exec – Decode a single tile
+- exec – Decode a single tile
-+ merge – Combine decoded tiles into the final output
+- merge – Combine decoded tiles into the final output
 #### Execution Flow
-+ Call split(z) to generate a list of TileTask and a GridSpec
+- Call split(z) to generate a list of TileTask and a GridSpec
-+ Dispatch tasks across ranks using workload-based balancing
+- Dispatch tasks across ranks using workload-based balancing
-+ Each rank executes exec(task) on its assigned tiles
+- Each rank executes exec(task) on its assigned tiles
-+ Gather decoded tile results to rank 0
+- Gather decoded tile results to rank 0
-+ Rank 0 performs merge(...)
+- Rank 0 performs merge(...)
-+ (Optional) Broadcast final result to all ranks
+- (Optional) Broadcast final result to all ranks
 This design separates:
-+ Distributed execution logic
+- Distributed execution logic
-+ Model-specific tiling and merging logic
+- Model-specific tiling and merging logic
 #### Why split / exec / merge is necessary?
@@ -82,23 +83,23 @@ The latent tensor cannot be arbitrarily partitioned.
 During decoding:
-+ Each output pixel may depend on neighboring pixels
+- Each output pixel may depend on neighboring pixels
-+ The receptive field is model-dependent
+- The receptive field is model-dependent
 Therefore:
-+ Tiles must include overlap
+- Tiles must include overlap
-+ Merge must perform blending to avoid seams
+- Merge must perform blending to avoid seams
 ## Step-by-Step Implementation (Decode)
 ### Step 1: Implement DistributedAutoencoderKLQwenImage
 `QwenImagePipeline` use `AutoencoderKLQwenImage` for vae, so implement a distributed version:
-
+```python
 ```
 class DistributedAutoencoderKLQwenImage(AutoencoderKLQwenImage, DistributedVaeMixin):
    @classmethod
    def from_pretrained(cls, *args: Any, **kwargs: Any):
@@ -106,28 +107,35 @@ class DistributedAutoencoderKLQwenImage(AutoencoderKLQwenImage, DistributedVaeMi
        model.init_distributed()
        return model
 ```
 **Key points**:
-+ Inherit both AutoencoderKLQwenImage and DistributedVaeMixin
+
-+ Call init_distributed() after loading weights
+- Inherit both AutoencoderKLQwenImage and DistributedVaeMixin
 - Call init_distributed() after loading weights
 ### Step 2: Implement split/exec/merge
 Reuse `AutoencoderKLQwenImage.tiled_decode` logic and divide it into three stages. And we need return tiles with `GridSpec` and `TileTask`:
-```
+
 ```python
 class GridSpec:
    split_dims: tuple[int, ...]  # Tensor dimensions being split (e.g., (2, 3) for (B, C, H, W))
    grid_shape: tuple[int, ...]  # Tile grid layout (num_rows, num_cols)
    tile_spec: dict = field(default_factory=dict) # Metadata required for merging
    output_dtype: torch.dtype | None = None # Final output dtype
 ```
-```
+
 ```python
 class TileTask:
    tile_id: int # task id
    grid_coord: tuple[int, ...]  # Tile position in grid
    tensor: torch.Tensor | list[torch.Tensor]  # The tile tensor
    workload: int | float = 1 # Used for load balancing (e.g., tile area)
 ```
 And tiled base split/exec/merge as follow:
-```
+
 ```python
 def tile_split(self, z: torch.Tensor) -> tuple[list[TileTask], GridSpec]:
    # mostly copy from AutoencoderKL
    _, _, num_frames, height, width = z.shape
@@ -210,11 +218,14 @@ def tile_merge(self, coord_tensor_map: dict[tuple[int, ...], torch.Tensor], grid
 ```
 ### Step 3: Override tiled_decode
 We need to override tiled_decode, the main logic is:
-+ check distributed is enabled
+
-+ select split/exec/merge
+- check distributed is enabled
-+ Invoke self.distributed_executor.execute to decode
+- select split/exec/merge
-```
+- Invoke self.distributed_executor.execute to decode
 ```python
 def tiled_decode(self, z: torch.Tensor, return_dict: bool = True):
    if not self.is_distributed_enabled():
        return super().tiled_decode(z, return_dict=return_dict)
@@ -230,11 +241,14 @@ def tiled_decode(self, z: torch.Tensor, return_dict: bool = True):
    return DecoderOutput(sample=result)
 ```
 `broadcast_result` is set to True or False depending on the model; when enabled, the result will be used even on ranks other than 0.
 ### Step 4: Modify Pipeline
 Change vae model from AutoencoderKLQwenImage to DistributedAutoencoderKLQwenImage
-```
+
 ```diff
 class YourModelPipeline(nn.Module):
    def __init__(
        self,
@@ -259,8 +273,8 @@ For models that require VAE encoding (e.g., Image-to-Video), you can also parall
 Similar to decode, split the input tensor into tiles. Key considerations:
-+ **Patchify handling**: If the model uses `patch_size`, scale tile parameters accordingly
+- **Patchify handling**: If the model uses `patch_size`, scale tile parameters accordingly
-+ **Temporal chunking**: Video VAEs may have temporal compression (e.g., 4x)
+- **Temporal chunking**: Video VAEs may have temporal compression (e.g., 4x)
 ```python
 def encode_tile_split(self, x: torch.Tensor) -> tuple[list[TileTask], GridSpec]:
@@ -406,23 +420,28 @@ def tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
 **Key differences from decode parallel:**
 | Aspect | Decode Parallel | Encode Parallel |
-|--------|-----------------|-----------------|
+| -------- | ----------------- | ----------------- |
 | `broadcast_result` | Often `False` (only rank 0 needs output) | `True` (all ranks need latents for diffusion) |
 | Patchify | Applied in merge (unpatchify) | Handled by parent `_encode()` before `tiled_encode()` |
 | Temporal chunking | Frame-by-frame | Chunk-based (e.g., 1 + 4n frames) |
 ## Testing
 Verify numerical consistency between:
 + vae_patch_parallel_size = 1
-+ vae_patch_parallel_size = N
+Verify numerical consistency between:
 - vae_patch_parallel_size = 1
 - vae_patch_parallel_size = N
 Example:
 ```python
 torch.allclose(output_1, output_n, atol=1e-5)
 ```
 Testing requirements:
-+ Fix random seed
+
-+ Use identical tiling strategy
+- Fix random seed
 - Use identical tiling strategy
 ```python
 m = Omni(
@@ -434,6 +453,7 @@ m = Omni(
        ),
    )
 ```
 When vae_patch_parallel_size is larger than the DiT world size, it will automatically fall back to using the DiT world size instead.
 ## Reference Implementations
@@ -441,7 +461,7 @@ When vae_patch_parallel_size is larger than the DiT world size, it will automati
 Complete examples in the codebase:
 | Model | Path | Decode Parallel | Encode Parallel |
-|-------|------|-----------------|-----------------|
+| ------- | ------ | ----------------- | ----------------- |
 | **Z-Image** | `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl.py` | ✅ | ❌ |
 | **Wan2.2** | `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_wan.py` | ✅ | ✅ |
 | **Qwen-Image** | `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py` | ✅ | ❌ |
@@ -456,7 +476,7 @@ The tile-parallel executor above assigns independent spatial **tiles** to ranks.
 ### How it differs from tile parallel
 | Aspect | Tile parallel (`"tile"`) | Spatially-sharded (`"spatial_shard_height"`/`"spatial_shard_width"`) |
-|--------|--------------------------|-----------------------------------------------|
+| -------- | -------------------------- | ----------------------------------------------- |
 | Unit of work | Independent overlapping tiles | A single global feature map sharded along H or W |
 | Cross-rank communication | Gather tiles to rank 0, stitch + blend | Per-conv **halo exchange** of boundary rows/cols (P2P) |
 | Output assembly | Blend overlapping tiles | All-gather shards on rank 0, trim padding (matches `broadcast_result=False`) |
@@ -481,7 +501,7 @@ serve.py (--vae-parallel-mode) / OmniEngineArgs
 ### Notes
 - The decoder is patched **in place** the first time spatial-shard decode runs and is bound to a single split dimension for the lifetime of the VAE instance.
- Numerical correctness vs. single-GPU decode is covered by `tests/diffusion/distributed/test_wan_spatial_shard.py::test_spatial_shard_decode_matches_reference` (multi-GPU, nightly `full_model` + `distributed_cuda`).
+- Numerical correctness vs. single-GPU decode is covered by `tests/diffusion/distributed/test_wan_spatial_shard.py::test_spatial_shard_decode_matches_reference` (multi-GPU, nightly `full_model` + `cards_2`).
 ---
@@ -37,6 +37,7 @@ validation_paths:
  - tests/engine/test_async_omni_engine_input.py
  - tests/engine/test_async_omni_engine_outputs.py
  - tests/engine/test_async_omni_engine_abort.py
  - tests/e2e/offline_inference/test_qwen3_omni.py
  - tests/engine/test_async_omni_engine_stage_init.py
  - tests/engine/test_orchestrator.py
  - tests/engine/test_orchestrator_error_handling.py
@@ -113,6 +114,15 @@ Numbers become append-only after normative promotion.
 Test routing, output ordering, queue correlation, cancellation, failure
 propagation, shutdown ordering, and representative multi-stage execution.
 AR abort of a final-stage LLM request must deliver a terminal output with
 `finish_reason="abort"` and the generated prefix so collocated training can
 resume; diffusion abort remains whole-sample retry. Aborting a parallel-sampling
 child must drop `parent_requests` once no children remain. EngineCore control
 RPCs (`sleep` / `wake_up` / `pause_scheduler` / `resume_scheduler`) must
 propagate worker exceptions rather than returning `{"supported": False}`.
 When a request has multiple final output stages, only the last abort message
 is request-terminal. Output-processor abort state is committed only after the
 physical EngineCore abort succeeds.
 Changes that cross into stage runtime or public error behavior require review
 from that contract's owners.
@@ -43,6 +43,7 @@ depends_on:
 validation_paths:
  - tests/entrypoints/test_omni_entrypoints.py
  - tests/entrypoints/test_async_omni.py
  - tests/entrypoints/test_async_omni_pause_sleep_routing.py
  - tests/entrypoints/test_async_omni_duplex.py
  - tests/entrypoints/test_serve.py
  - tests/entrypoints/test_stream_finish_reason.py
@@ -119,7 +120,13 @@ promotion.
 Test request validation, protocol conversion, model-adapter routing,
 streaming, session identity, disconnect/cancellation, and error mapping for
-each affected offline or serving entrypoint.
+each affected offline or serving entrypoint. Sleep must wait for in-flight
 `generate()` admission before EngineCore offload; `wake_up` does not resume
 admission — callers must `resume_generation()`. Sleeping tags are tracked per
 stage so `wake_up(stage_ids=[0])` does not skip a later `wake_up(stage_ids=[1])`.
 Streaming input pumps take an admission slot immediately before each EngineCore
 ADD or update, not while waiting for the next client chunk. Frontend abort
 keeps `request_states` until `generate()` consumes the terminal output.
 ## Promotion gate
@@ -257,6 +257,53 @@ Lookup checks the marker both before validation and after constructing the
 lease so a weaker overlapping lookup cannot return a lease after a stronger
 validation has denied the artifact.
 If publishing the external deny marker fails, lookup returns a retryable
 storage failure and marks the artifact directory with a persistent invalidation
 mode before releasing the shared lock. Cooperating readers check that mode both
 while opening and immediately before returning a lease. Lookup then makes a
 bounded attempt to take the exclusive artifact lock and quarantine the entry;
 the quarantine transition removes the internal mode marker. If a cooperating
 builder replaces the marked inode while the locks are exchanged, containment
 preserves that new publication instead of quarantining it. If neither the
 external marker nor the directory-mode marker can be persisted, cross-process
 exclusion cannot be guaranteed on an unavailable filesystem and the domain
 must remain failed rather than treating the artifact as an ordinary invalid
 cache entry.
 New publications carry the internal invalidation mode through the staging
 rename. Hardening clears it and is the logical authority transition for a
 process crash; the temporary- and artifact-parent syncs then make the rename
 durable, with an artifact-inode sync first making the hardened mode durable. A
 process that dies before hardening leaves a persistently denied
 entry, while one that dies after successful hardening leaves an authoritative
 ready entry. Any synchronous failure after rename is marked invalid and
 contained in quarantine before `PUBLICATION_FAILED` is returned, so that
 nonretryable result cannot later resolve as a ready hit. If the filesystem
 refuses both the inode marker and the quarantine move, the store instead returns
 a retryable storage failure with `outcome_uncertain` details; as with deny-marker
 double failure, cross-process exclusion cannot be guaranteed until storage
 recovers. Once the artifact-parent sync succeeds, a later artifact- or
 build-lock release error is logged and lookup resolves the committed entry
 instead of reporting a contradictory publication failure. The same authority
 rule applies when a competing publisher already installed byte-identical
 semantic content; staging-cleanup errors are retried without downgrading that
 validated ready entry to publication failure.
 Before widening a hardened artifact for a lifecycle move, the store durably
 sets the same invalidation mode. A process crash before rename therefore leaves
 the writable source denied. After rename, the store re-hardens the destination
 but retains the marker until both parent-directory syncs make the move durable;
 only then does it durably clear the marker. An interrupted move therefore leaves
 a marked transition that the next exact-key lifecycle operation re-hardens.
 Explicit cleanup uses the locked move and retries exact-key
 `.cleanup.*` tombstones left by an interrupted removal; a later build performs
 the same tombstone reconciliation before producing replacement content.
 An operational failure while inspecting a noncooperative competing publication
 is a retryable storage failure. The competing entry remains authoritative and
 is not mislabeled as corrupt or quarantined merely because its manifest could
 not be read transiently.
 `inspect_domain()` and `inspect_artifact()` return typed, read-only snapshots.
 Inspection validation does not create a deny marker. Reported lock activity is
 point-in-time evidence only; kernel locks are not a persistent owner registry.
@@ -3,4 +3,6 @@
 vLLM-Omni's examples are split into two categories:
 - If you are using vLLM-Omni from within Python code, see the *Offline Inference* section.
- If you are using vLLM-Omni from an HTTP application or client, see the *Online Serving* section.
+- If you are using vLLM-Omni from an HTTP application or client, start with
  the [API Server guide](../serving/README.md) to choose an endpoint, then use
  the *Online Serving* section for model-specific examples.
@@ -23,6 +23,19 @@ vLLM-Omni currently recommends using the Docker image setup steps below.
 DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.xpu -t vllm-omni-xpu --shm-size=4g .
 ```
 This layers vLLM-Omni on top of the published `vllm/vllm-openai-xpu:<VLLM_VERSION>`
 base image, which Docker pulls automatically. To target a different vLLM release,
 pass `--build-arg VLLM_VERSION=<tag>`. If that tag has not been published yet, build
 the base from upstream's own Dockerfile first and point the build at it:
 ```bash
 git clone --depth 1 --branch "$VLLM_VERSION" https://github.com/vllm-project/vllm /tmp/vllm
 DOCKER_BUILDKIT=1 docker build -t vllm-openai-xpu:local --target vllm-openai \
  -f /tmp/vllm/docker/Dockerfile.xpu /tmp/vllm
 DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.xpu -t vllm-omni-xpu --shm-size=4g \
  --build-arg VLLM_BASE=vllm-openai-xpu:local .
 ```
 #### Launch the docker image
 ##### Launch with OpenAI API Server
@@ -91,7 +91,7 @@ if __name__ == "__main__":
 For more usages, please refer to [offline inference](../user_guide/examples/offline_inference/qwen2_5_omni.md)
-## Online Serving with OpenAI-Completions API
+## Online Serving with the API Server
 Text-to-image generation quickstart with vLLM-Omni:
@@ -100,20 +100,16 @@ vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8091
 ```
 ```bash
-curl -s http://localhost:8091/v1/chat/completions \
+curl -s http://localhost:8091/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
-    "messages": [
+    "prompt": "a cup of coffee on the table",
-      {"role": "user", "content": "a cup of coffee on the table"}
+    "size": "1024x1024",
-    ],
+    "response_format": "b64_json",
-    "extra_body": {
+    "seed": 42
-      "height": 1024,
+  }' | jq -r '.data[0].b64_json' | base64 -d > coffee.png
      "width": 1024,
      "num_inference_steps": 50,
      "guidance_scale": 4.0,
      "seed": 42
    }
  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2 | base64 -d > coffee.png
 ```
-For more details, please refer to [online serving](../user_guide/examples/online_serving/text_to_image.md).
+See the [API Server guide](../serving/README.md) to choose an endpoint. For
 model-specific details, refer to the [text-to-image online serving
 example](../user_guide/examples/online_serving/text_to_image.md).
@@ -21,20 +21,20 @@ th {
 |--------------|--------|-------------------|------------|---------|-----|-----------|--------|
 | `Qwen3OmniMoeForConditionalGeneration` | Qwen3-Omni | `Qwen/Qwen3-Omni-30B-A3B-Instruct` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen3-Omni.md) |
 | `Qwen2_5OmniForConditionalGeneration` | Qwen2.5-Omni | `Qwen/Qwen2.5-Omni-7B`, `Qwen/Qwen2.5-Omni-3B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `MingFlashOmniForConditionalGeneration` + `MingImagePipeline` | Ming-flash-omni-2.0 (omni-speech + imagegen<sup>1</sup>) | `Jonathan1909/Ming-flash-omni-2.0` | ✅︎ |   |   |   | — |
+| `MingFlashOmniForConditionalGeneration` + `MingImagePipeline` | Ming-flash-omni-2.0 (omni-speech + imagegen<sup>1</sup>) | `Jonathan1909/Ming-flash-omni-2.0` | ✅︎ | | | | — |
 | `BagelForConditionalGeneration` | BAGEL (DiT-only) | `ByteDance-Seed/BAGEL-7B-MoT` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `InternVLAA1Pipeline` | InternVLA-A1 | `InternRobotics/InternVLA-A1-3B` | ✅︎ | ✅︎ | | | — |
 | `Gr00tN1d7Pipeline` | GR00T N1.7 | `nvidia/GR00T-N1.7-3B` | ✅︎ | | | | — |
 | `HunyuanImage3ForCausalMM` | HunyuanImage3.0 (DiT-only) | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` | ✅︎ | | | | [Published](https://recipes.vllm.ai/Qwen/Qwen-Image) |
+| `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` | ✅︎ | | | ✅︎ | [Published](https://recipes.vllm.ai/Qwen/Qwen-Image) |
-| `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen-Image-2512.md) |
+| `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` | ✅︎ | | | ✅︎ | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen-Image-2512.md) |
 | `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2511 | `Qwen/Qwen-Image-Edit-2511` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `GlmImagePipeline` | GLM-Image | `zai-org/GLM-Image` | ✅︎ | ✅︎ | | | — |
 | `ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `Krea2Pipeline` | Krea 2 (Raw + Turbo) | `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | ✅︎ | | | | — |
+| `Krea2Pipeline` | Krea 2 (Raw + Turbo) | `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | ✅︎ | | | ✅︎ | — |
 | `WanPipeline` | Wan2.1-T2V, Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`, `Wan-AI/Wan2.1-T2V-14B-Diffusers`, `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LingBotWorldCausalDMDPipeline` | LingBot-World 2.0 v1 (experimental) | `robbyant/lingbot-world-v2-14b-causal-fast-diffusers` | Experimental | | | | — |
@@ -42,7 +42,7 @@ th {
 | `SanaWmPipeline` | SANA-WM | `BBBBruce/SANA-WM_bidirectional-stage1-diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-WM.md) |
 | `Wan22S2VPipeline` | Wan2.2-S2V | `Wan-AI/Wan2.2-S2V-14B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22VACEPipeline` | Wan2.1-VACE | `Wan-AI/Wan2.1-VACE-1.3B-diffusers`, `Wan-AI/Wan2.1-VACE-14B-diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `Wan22VACEPipeline` | Wan2.2-VACE | `Pyros13/Wan2.2-VACE-Fun-A14B-Diffusers` | ✅︎ |   |   |   | — |
+| `Wan22VACEPipeline` | Wan2.2-VACE | `Pyros13/Wan2.2-VACE-Fun-A14B-Diffusers` | ✅︎ |   |   | ✅︎ | — |
 | `LTX2Pipeline` | LTX-2 / LTX-2.3 one-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | | — |
 | `LTX2TwoStagePipeline` | LTX-2 / LTX-2.3 ordinary two-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` + matching Lightricks LoRA and upsampler | ✅︎ | ✅︎ | | | — |
 | `LTX2DistilledOneStagePipeline` | LTX-2 / LTX-2.3 merged-distilled one-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` | ✅︎ | ✅︎ | | | — |
@@ -62,12 +62,12 @@ th {
 | `LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatVideoAvatarPipeline` | LongCat-Video-Avatar-1.5 A2V/AI2V (native single-speaker and multi-speaker AI2V/AVC) | `meituan-longcat/LongCat-Video-Avatar-1.5` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/meituan-longcat/LongCat-Video-Avatar-1.5.md) |
-| `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit` | ✅︎ | | | | — |
+| `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit`, `Boogu/Boogu-Image-0.1-Edit-Turbo` | ✅︎ | | | ✅︎ | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Boogu/Boogu-Image.md) |
 | `StableDiffusionXLPipeline` | Stable-Diffusion-XL | `stabilityai/stable-diffusion-xl-base-1.0` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ | — |
-| `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | | — |
+| `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | ✅︎ | — |
-| `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | | — |
+| `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` | ✅︎ | | | | — |
@@ -78,10 +78,10 @@ th {
 | `OmniGen2Pipeline` | OmniGen2 | `OmniGen2/OmniGen2` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `StableAudioPipeline` | Stable-Audio-Open | `stabilityai/stable-audio-open-1.0` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `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-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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-1.7B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `MingTTSForConditionalGeneration` | Ming-omni-tts dense 0.5B | `inclusionAI/Ming-omni-tts-0.5B` | ✅︎ | | | | — |
+| `MingTTSForConditionalGeneration` | Ming-omni-tts dense 0.5B | `inclusionAI/Ming-omni-tts-0.5B` | ✅︎ | | | ✅︎ | — |
 | `GLMTTSForConditionalGeneration` | GLM-TTS | `zai-org/GLM-TTS` | ✅︎ | | | | — |
 | `MossTTSNanoForCausalLM` | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | ✅︎ | | | | — |
 | `MossTTSDelayModel` | MOSS-TTS, MOSS-TTSD, MOSS-SoundEffect, MOSS-VoiceGenerator | `OpenMOSS-Team/MOSS-VoiceGenerator` | ✅︎ | | | | — |
@@ -101,10 +101,10 @@ th {
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | | — |
 | `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | | — |
 | `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | | — |
-|`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | | — |
+|`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | ✅︎ | — |
 | `MiniCPMO45OmniForConditionalGeneration` | MiniCPM-o 4.5 | `openbmb/MiniCPM-o-4_5` | ✅︎ | | ✅︎ | | — |
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `GepardTalkerForConditionalGeneration` | Gepard-1.0 | `nineninesix/gepard-1.0` | ✅︎ | | | | — |
+| `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` | ✅︎ | | | | — |
@@ -0,0 +1,187 @@
 # API Server
 vLLM-Omni exposes OpenAI-compatible HTTP APIs plus vLLM-Omni extensions for
 image, audio, video, realtime, and robot-policy workloads. Use this page to
 choose an endpoint; follow the linked reference page for model-specific fields,
 wire protocols, and response details.
 Each server instance hosts one model. An endpoint is usable only when the
 loaded model supports that task.
 This guide focuses on vLLM-Omni's primary task endpoints. Compatible models
 may also expose standard routes inherited from vLLM, such as
 `POST /v1/completions`; those routes remain model- and task-dependent.
 ## Start and Verify the Server
 Start a model with the shared serving command:
 ```bash
 vllm serve <model> --omni --port 8091
 ```
 Some models require additional flags or a deployment configuration. Use the
 model-specific guide when one is provided.
 After the server starts, check its health and served model name:
 ```bash
 export VLLM_OMNI_BASE_URL=http://localhost:8091
 curl "$VLLM_OMNI_BASE_URL/health"
 curl "$VLLM_OMNI_BASE_URL/v1/models" | jq .
 ```
 Use `http://localhost:8091/v1` as the `base_url` for an OpenAI SDK client. If
 the server was started with `--api-key`, include an `Authorization` header with
 `Bearer <api-key>` in requests.
 ## Choose a Core HTTP Endpoint
 Prefer a task-specific endpoint when one matches your workload. Use Chat
 Completions for conversational or heterogeneous omni pipelines, rather than as
 the default wrapper for every generation task.
 | Task | Endpoint | Request | Response | Details |
 |------|----------|---------|----------|---------|
 | Conversation or multimodal understanding/generation | `POST /v1/chat/completions` | JSON | OpenAI-style JSON or SSE | [Chat Completions](chat_completions_api.md) |
 | Text-to-speech | `POST /v1/audio/speech` | JSON | Audio bytes or SSE | [Speech](speech_api.md) |
 | Sound, music, or ambient audio generation | `POST /v1/audio/generate` | JSON | Audio bytes | [Audio Generation](audio_generate_api.md) |
 | Text-to-image | `POST /v1/images/generations` | JSON | Base64 JSON or an image file | [Image Generation](image_generation_api.md) |
 | Image editing | `POST /v1/images/edits` | Multipart form | Base64 JSON or SSE | [Image Edit](image_edit_api.md) |
 | Video generation (recommended) | `POST /v1/videos` | Multipart form | Asynchronous job JSON | [Videos](videos_api.md) |
 | Video generation (blocking) | `POST /v1/videos/sync` | Multipart form | Video bytes | [Videos](videos_api.md#synchronous-response) |
 `POST /v1/videos/sync` is intended for tests and simple scripts. Use the
 asynchronous job API for production requests so clients can poll, download,
 and delete generated videos independently.
 ## Minimal Requests
 The examples below assume a compatible model is already running. TTS voices,
 media inputs, generation controls, and output capabilities vary by model.
 === "Chat"
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/chat/completions" \
      -H "Content-Type: application/json" \
      -d '{
        "messages": [{"role": "user", "content": "Describe vLLM-Omni briefly."}],
        "modalities": ["text"]
      }'
    ```
 === "Speech"
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/audio/speech" \
      -H "Content-Type: application/json" \
      -d '{
        "input": "Hello from vLLM-Omni.",
        "voice": "vivian"
      }' \
      --output speech.wav
    ```
 === "Audio"
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/audio/generate" \
      -H "Content-Type: application/json" \
      -d '{
        "input": "Ocean waves on a quiet beach",
        "audio_length": 5.0
      }' \
      --output audio.wav
    ```
 === "Image"
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/images/generations" \
      -H "Content-Type: application/json" \
      -d '{
        "prompt": "A small robot reading beside a window",
        "size": "1024x1024",
        "response_format": "file"
      }' \
      --output image.png
    ```
 === "Image edit"
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/images/edits" \
      -F "image=@input.png" \
      -F "prompt=Turn the background into a snowy mountain" \
      | jq -r '.data[0].b64_json' \
      | base64 --decode > edited.png
    ```
 === "Video"
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/videos" \
      -F "prompt=A cinematic tracking shot of a mountain lake at sunrise" \
      -F "size=1280x720" \
      -F "seconds=5"
    ```
    The response contains a video job `id`. Poll and download it with:
    ```bash
    curl "$VLLM_OMNI_BASE_URL/v1/videos/<video_id>"
    curl -L "$VLLM_OMNI_BASE_URL/v1/videos/<video_id>/content" --output video.mp4
    ```
 ## Choose a Streaming or Realtime Endpoint
 These WebSocket APIs have different event schemas and cannot be used
 interchangeably.
 | Workload | Endpoint | Interaction model | Details |
 |----------|----------|-------------------|---------|
 | Incremental text input for speech synthesis | `WS /v1/audio/speech/stream` | Send text events and receive audio | [Streaming Text to Speech](speech_api.md#streaming-text-input-websocket) |
 | Live video understanding | `WS /v1/video/chat/stream` | Send video frames and receive text/audio | [Streaming Video Input](video_stream_api.md) |
 | Turn-based realtime audio | `WS /v1/realtime` | Stream one audio input and receive transcript/audio events | [Realtime Audio](realtime_api.md) |
 | Continuous speech-to-speech interaction | `WS /v1/realtime?duplex=1` or `WS /v1/duplex` | Listen and speak concurrently with session control | [Full Duplex](full_duplex_api.md) |
 | Generated video chunks | `WS /v1/realtime/video` | Start a diffusion request and receive fragmented MP4 | [Streaming Video Output](streaming_video_output_api.md) |
 | Robot policy inference | `WS /v1/realtime/robot/openpi` | Send MessagePack observations and receive action arrays | [OpenPI Robot Policy](openpi_api.md) |
 All six routes are model- or configuration-dependent. In particular,
 `/v1/realtime` is not full duplex unless the client sets `duplex=1` and the
 deployment explicitly enables duplex sessions. Clients should also verify the
 duplex capability payload because the query-parameter form falls back to the
 ordinary realtime handler when duplex is unavailable.
 ## Related Endpoints
 | Purpose | Endpoints | Reference |
 |---------|-----------|-----------|
 | Discovery and readiness | `GET /health`, `GET /v1/models` | This page |
 | Batched conversations | `POST /v1/chat/completions/batch` | [Batch requests](chat_completions_api.md#batch-requests) |
 | Batched speech | `POST /v1/audio/speech/batch` | [Batch speech generation](speech_api.md#batch-speech-generation) |
 | TTS voice management | `GET/POST /v1/audio/voices`, `DELETE /v1/audio/voices/{name}` | [Voices](speech_api.md#voices-endpoint) |
 | Video job lifecycle | `GET /v1/videos`, `GET/DELETE /v1/videos/{video_id}`, `GET /v1/videos/{video_id}/content` | [Video endpoints](videos_api.md#endpoints) |
 | Release and restore stage memory | `POST /v1/omni/sleep`, `POST /v1/omni/wakeup` | [Sleep Mode](../features/sleep_mode.md) |
 ## Standalone Experimental Servers
 PersonaPlex also provides a Moshi-compatible WebSocket server, while JoyVL
 provides a stateful interaction orchestrator in front of another model server.
 These are separate processes with their own routes; they are not additional
 paths on every `vllm serve --omni` instance. See [Standalone Experimental
 Servers](standalone_servers.md) before deploying either one.
 ## Compatibility Notes
 - A registered path does not mean every model supports it. Unsupported tasks
  return an OpenAI-style error or a WebSocket `error` event.
 - Use JSON for Chat, Speech, Audio Generation, and Image Generation. Use
  `multipart/form-data` when uploading images, videos, or other files.
 - Generation can take minutes for large diffusion models. Configure client
  timeouts accordingly; prefer asynchronous video jobs over long blocking
  requests.
 - Model-specific parameters are vLLM-Omni extensions. Check the endpoint
  reference before assuming an OpenAI SDK exposes them directly.
@@ -4,14 +4,14 @@ vLLM-Omni provides an API for text-to-audio generation using diffusion-based mod
 Unlike the [Speech API](speech_api.md) which targets text-to-speech synthesis, the Audio Generate API is designed for general-purpose audio generation from text descriptions (sound effects, music, ambient soundscapes, etc.).
-Each server instance runs a single model (specified at startup via `vllm-omni serve <model> --omni`).
+Each server instance runs a single model (specified at startup via `vllm serve <model> --omni`).
 ## Quick Start
 ### Start the Server
 ```bash
-vllm-omni serve stabilityai/stable-audio-open-1.0 \
+vllm serve stabilityai/stable-audio-open-1.0 \
    --host 0.0.0.0 \
    --port 8091 \
    --gpu-memory-utilization 0.9 \
@@ -326,7 +326,7 @@ curl http://localhost:8091/health
 Enable debug logging:
 ```bash
-vllm-omni serve stabilityai/stable-audio-open-1.0 \
+vllm serve stabilityai/stable-audio-open-1.0 \
    --host 0.0.0.0 \
    --port 8091 \
    --gpu-memory-utilization 0.9 \
@@ -1,78 +1,110 @@
 # Chat Completions API
-vLLM-Omni supports generating and editing images via the `/v1/chat/completions`
+Use `/v1/chat/completions` for conversational and multimodal pipelines that
-endpoint using diffusion models. This page explains how to pass generation
+follow the OpenAI Chat API schema. Depending on the loaded model, a request can
-parameters (such as `num_inference_steps`, `height`, `width`) to diffusion
+contain text, image, audio, or video input and can return text, audio, images,
-models through this endpoint.
+or other model-specific output.
-!!! tip
+Some diffusion pipelines also support image generation and editing through
-    For dedicated endpoints that accept generation parameters as top-level
+Chat Completions. Prefer the dedicated [Image Generation
-    fields, see [Image Generation API](image_generation_api.md) and
+API](image_generation_api.md) or [Image Edit API](image_edit_api.md) for
-    [Image Edit API](image_edit_api.md).
+task-specific clients; their request fields and response contracts are more
 direct.
-## Passing Generation Parameters
+## Basic Request
 The `/v1/chat/completions` endpoint follows the OpenAI Chat API schema, which
 does not natively include diffusion-specific fields like `num_inference_steps`
 or `height`. How you pass these extra fields depends on your client.
 ### curl / Python `requests`
 Wrap generation parameters inside an `"extra_body"` key in the JSON body:
 ```bash
-curl -s http://localhost:8091/v1/chat/completions \
+curl http://localhost:8091/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
-      {"role": "user", "content": "A beautiful landscape painting"}
+      {"role": "user", "content": "Describe vLLM-Omni briefly."}
    ],
-    "extra_body": {
+    "modalities": ["text"]
      "num_inference_steps": 50,
      "seed": 42
    }
  }'
 ```
-### OpenAI Python SDK
+Use `stream: true` for Server-Sent Events. Input media syntax, supported output
 `modalities`, and response choices depend on the model; see the
 [model-specific online serving examples](../user_guide/examples/online_serving/qwen3_omni.md).
-Use the `extra_body` **keyword argument**. The SDK automatically merges these
+!!! tip
-fields into the top-level request body:
+    Each server hosts one model. Query `GET /v1/models` for its served model
    name when your client requires the `model` field.
-```python
+## vLLM Extension Parameters
-response = client.chat.completions.create(
+
-    model="Qwen/Qwen-Image",
+Send standard and vLLM-specific Chat parameters as top-level fields in a
-    messages=[{"role": "user", "content": "A beautiful landscape painting"}],
+direct HTTP request. When using the OpenAI Python SDK, pass vLLM-specific
-    extra_body={
+fields through the SDK's `extra_body` keyword; the SDK merges them into the
-        "num_inference_steps": 50,
+top-level JSON sent to the server.
-        "seed": 42,
+
-    },
+=== "curl"
-)
+
    ```bash
    curl http://localhost:8091/v1/chat/completions \
      -H "Content-Type: application/json" \
      -d '{
        "messages": [{"role": "user", "content": "Write a short haiku."}],
        "top_k": 40
      }'
    ```
 === "OpenAI Python SDK"
    ```python
    from openai import OpenAI
    client = OpenAI(base_url="http://localhost:8091/v1", api_key="none")
    response = client.chat.completions.create(
        model="your-served-model",
        messages=[{"role": "user", "content": "Write a short haiku."}],
        extra_body={"top_k": 40},
    )
    ```
 ??? note "Diffusion compatibility"
    Diffusion models exposed through Chat Completions can accept supported
    fields such as `num_inference_steps`, `seed`, `height`, and `width`. Send
    them at the top level with direct HTTP, or in the SDK's `extra_body`
    keyword argument.
    A literal nested `"extra_body"` JSON object is accepted for compatibility,
    but it is not recommended for new direct HTTP clients. Do not provide the
    same parameter in multiple locations; the server returns a `400` error for
    duplicate diffusion parameters. Prefer the dedicated [Image Generation
    API](image_generation_api.md) or [Image Edit API](image_edit_api.md) when
    either endpoint matches the task.
 ## Batch Requests
 `POST /v1/chat/completions/batch` accepts the same shared generation fields,
 but `messages` is a list of conversations. The response contains one choice
 per conversation in input order.
 ```bash
 curl http://localhost:8091/v1/chat/completions/batch \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      [{"role": "user", "content": "Summarize vLLM in one sentence."}],
      [{"role": "user", "content": "Summarize vLLM-Omni in one sentence."}]
    ],
    "max_tokens": 64
  }'
 ```
-!!! note "SDK `extra_body` vs. JSON `extra_body`"
+Batch Chat Completions does not support streaming, tools, beam search, or
-    These two `extra_body` usages look similar but work differently under the
+`n > 1`.
    hood. The SDK flattens the dict into the top-level request JSON, while the
    curl/requests approach sends it as a nested `"extra_body"` key. Both are
    handled correctly by the server.
 !!! note "About the `ignored fields` warning"
    You may see a log message like:
    ```
    WARNING: The following fields were present in the request but ignored: {'height', 'width', ...}
    ```
    This is **harmless**. It is emitted by vLLM's request validation layer
    because these fields are not part of the standard OpenAI
    `ChatCompletionRequest` schema. The fields are still stored internally
    and correctly forwarded to the diffusion pipeline.
 ## Model-Specific Examples
-For complete examples with full request/response details, see the model-specific
+For complete examples with model-specific inputs and outputs, see:
 guides:
 - [Qwen3-Omni](../user_guide/examples/online_serving/qwen3_omni.md)
 - [Qwen2.5-Omni](../user_guide/examples/online_serving/qwen2_5_omni.md)
 - [Text-to-Image (Qwen-Image)](../user_guide/examples/online_serving/text_to_image.md)
 - [Image-to-Image (Qwen-Image-Edit, Qwen-Image-Layered)](../user_guide/examples/online_serving/image_to_image.md)
 - [GLM-Image](../user_guide/examples/online_serving/glm_image.md)
@@ -0,0 +1,116 @@
 # Full-Duplex WebSocket API
 vLLM-Omni provides an experimental full-duplex runtime for models that can
 continue receiving speech while producing speech. It adds persistent session
 state, model-specific turn policy, overlap handling, playback acknowledgement,
 and optional session resume.
 Full duplex is distinct from the turn-based [Realtime Audio API](realtime_api.md).
 ## Choose an Endpoint
 | Endpoint | Protocol | Recommended use |
 |----------|----------|-----------------|
 | `WS /v1/realtime?duplex=1` | OpenAI Realtime-style event projection | Applications and browser clients |
 | `WS /v1/duplex` | Native vLLM-Omni duplex events | Runtime integration and low-level testing |
 Both endpoints use the same duplex engine and require the same model-side
 adapter. Prefer the Realtime projection unless the native lifecycle events are
 specifically required.
 ## Enable Full Duplex
 The route becomes usable only when the deployment configuration explicitly
 sets:
 ```yaml
 session_mode: duplex
 ```
 The selected model pipeline must also provide a duplex serving adapter.
 Model-native deployments configure the corresponding engine runtime extension
 and control plane as part of their registered pipeline.
 !!! warning
    `WS /v1/duplex` fails with `Duplex API is not available` when duplex is not
    enabled. By contrast, `/v1/realtime?duplex=1` falls back to the ordinary
    turn-based realtime handler when no duplex handler exists. Confirm that
    `session.created.session.capabilities` is present before treating the
    connection as full duplex.
 The current unified-runtime integrations are:
 - MiniCPM-o 4.5, using `vllm_omni/deploy/minicpmo_4_5_duplex.yaml`;
 - PersonaPlex, whose default `vllm_omni/deploy/personaplex.yaml` enables duplex.
 JoyVL is a separate HTTP interaction orchestrator and does not use these
 WebSocket endpoints. See [Standalone Experimental Servers](standalone_servers.md).
 ## MiniCPM-o Quick Start
 Start the duplex deployment:
 ```bash
 vllm serve openbmb/MiniCPM-o-4_5 --omni \
  --deploy-config vllm_omni/deploy/minicpmo_4_5_duplex.yaml \
  --trust-remote-code \
  --port 8091
 ```
 Stream a mono, PCM16, 16 kHz WAV file with the provided client:
 ```bash
 python examples/online_serving/minicpmo/realtime_duplex_demo.py \
  --url 'ws://localhost:8091/v1/realtime?duplex=1' \
  --model openbmb/MiniCPM-o-4_5 \
  --input-wav input_16k_mono.wav \
  --ref-audio reference_voice.wav \
  --output-dir /tmp/minicpmo-duplex
 ```
 ## Realtime Event Lifecycle
 A typical `/v1/realtime?duplex=1` session follows this lifecycle:
 1. Send `session.update` with the model, modalities, audio formats, and session
   options.
 2. Wait for `session.created`; inspect `session.capabilities` instead of
   assuming every model supports the same controls.
 3. Send `input_audio_buffer.append` events while microphone audio arrives.
 4. Send `input_audio_buffer.commit` at a user-turn boundary when required by
   the model policy.
 5. Consume `response.created`, transcript deltas, `response.audio.delta`, and
   `response.done` or `response.listen` events.
 6. Send `playback.ack` after audio has been played when the session advertises
   playback acknowledgement support.
 7. Send `session.close` and wait for `session.closed`.
 Unlike the turn-based realtime endpoint, input may continue while a response
 is active. The server can emit `overlap.decision` to describe whether input was
 deferred, treated as a short acknowledgement, or used to interrupt output.
 ## Capabilities and Model Differences
 The `session.created` payload includes capability fields such as
 `supports_barge_in`, `supports_playback_ack`, `supports_multi_session`,
 `supports_session_resume`, and `chunk_period_ms`. Treat this payload as the
 runtime contract.
 For example, PersonaPlex supports native overlapping speech but currently
 advertises `supports_barge_in=false`; destructive output interruption and
 model-state rewind have not been validated for that integration. Capacity and
 session-resume behavior also depend on the selected deployment configuration.
 ## Native Protocol
 `WS /v1/duplex` exposes lower-level lifecycle names including
 `session.create`, `input_audio_buffer.append`, `turn.signal`, `playback.ack`,
 and `session.close`. It returns native session, input, response, overlap, and
 error events. This protocol is experimental and may evolve with the runtime;
 applications should use the provided Realtime client where possible.
 See the [MiniCPM-o example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/minicpmo),
 [PersonaPlex example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/personaplex),
 and [full-duplex runtime design](https://github.com/vllm-project/vllm-omni/blob/main/vllm_omni/experimental/fullduplex/DESIGN.md)
 for model-specific validation and architecture details.
@@ -0,0 +1,109 @@
 # OpenPI Robot Policy WebSocket API
 Use `WS /v1/realtime/robot/openpi` for low-latency robot-policy inference with
 OpenPI-compatible clients. A persistent WebSocket carries observations to the
 loaded policy model and returns action arrays.
 Despite the `/realtime` prefix, this endpoint does not use the OpenAI Realtime
 event schema. It uses binary MessagePack frames with NumPy extensions.
 ## Availability
 The endpoint is available only when the loaded diffusion policy configuration
 contains `policy_server_config`. The server sends that model-specific
 configuration to the client as the first binary frame.
 For example, DreamZero declares:
 ```yaml
 policy_server_config:
  image_resolution: [180, 320]
  n_external_cameras: 2
  needs_wrist_camera: true
  needs_stereo_camera: false
  needs_session_id: true
  action_space: joint_position
 ```
 Without this configuration, the WebSocket returns `Robot policy not available`
 and closes.
 ## DreamZero Quick Start
 Start the bundled DreamZero deployment:
 ```bash
 vllm serve GEAR-Dreams/DreamZero-DROID --omni --port 8091 \
  --served-model-name dreamzero-droid \
  --deploy-config vllm_omni/deploy/dreamzero_tp1_cfg2.yaml \
  --enforce-eager --disable-log-stats
 ```
 Install the optional client dependencies, download the sample camera inputs,
 and run the client:
 ```bash
 pip install openpi-client websockets opencv-python
 hf download YangshenDeng/vllm-omni-dreamzero-assets \
  --repo-type dataset \
  --local-dir outputs/dreamzero/assets
 python examples/online_serving/dreamzero/openpi_client.py \
  --host 127.0.0.1 \
  --port 8091 \
  --video-dir outputs/dreamzero/assets
 ```
 ## Protocol
 The connection is request-response after an initial server handshake:
 ```text
 connect
  <- msgpack(policy_server_config)
 infer
  -> msgpack({"endpoint": "infer", "session_id": "...", ...observation})
  <- msgpack(ndarray | dict[str, ndarray])
 reset
  -> msgpack({"endpoint": "reset"})
  <- msgpack({"status": "reset successful"})
 ```
 If `endpoint` is omitted, the server treats the message as `infer`. Each
 inference message produces one action response after the engine request
 completes; action tokens or intermediate tensors are not streamed.
 NumPy arrays use the marker format implemented by `openpi-client`. The server
 also accepts the legacy vLLM NumPy marker representation on input. JSON text
 frames are not observation messages.
 ## Sessions and Reset
 - The API layer's current-session and first-call counters are scoped to each
  WebSocket connection.
 - `session_id` identifies model-side state across observations on that
  connection. If omitted, it defaults to `default`.
 - The first inference for a session is sent to the model with `reset=true`.
 - Changing `session_id`, or sending the `reset` command, causes the next
  inference to start with `reset=true`.
 - The policy pipeline owns observation transforms and persistent model state,
  normally keyed by `session_id`; the API layer forwards the raw observation
  dictionary.
 ## Limits and Errors
 - Maximum inbound payload size is 64 MiB.
 - The server closes an idle connection after 30 seconds.
 - Invalid binary input returns a MessagePack
  `{"type":"error","message":"Invalid request payload"}` response.
 - Inference failures return a generic `Internal inference error` without
  exposing an internal traceback.
 - Unsupported NumPy object, structured, and complex dtypes are rejected.
 Observation keys, camera layout, state tensors, and action shapes are defined
 by the loaded policy rather than by this transport. Read the handshake before
 constructing observations. See the [DreamZero example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/dreamzero)
 for a complete OpenPI client and DROID simulation loop.
@@ -0,0 +1,82 @@
 # Realtime Audio WebSocket API
 Use `WS /v1/realtime` to stream PCM audio into a compatible omni model and
 receive incremental transcription and synthesized response audio. This is a
 turn-based streaming API: it starts one generation over an audio stream and
 finishes after the client closes that input stream.
 For a model that listens while it is already speaking, use the experimental
 [Full-Duplex API](full_duplex_api.md) instead.
 ## Quick Start
 Start a compatible model. The currently documented example uses Qwen3-Omni:
 ```bash
 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 Run the provided client with a mono, 16-bit PCM, 16 kHz WAV file:
 ```bash
 python examples/online_serving/qwen3_omni/openai_realtime_client.py \
  --url ws://localhost:8091/v1/realtime \
  --model Qwen/Qwen3-Omni-30B-A3B-Instruct \
  --input-wav input_16k_mono.wav \
  --output-wav response.wav \
  --output-text response.txt
 ```
 The client dependency is `websockets`.
 ## Protocol
 Messages in both directions are JSON text frames. Audio carried inside an
 event is base64-encoded raw PCM16.
 | Direction | Event | Purpose |
 |-----------|-------|---------|
 | Server to client | `session.created` | Confirms that the WebSocket connection is ready |
 | Client to server | `session.update` | Selects and validates the served model |
 | Client to server | `input_audio_buffer.commit` with `final: false` | Starts generation over the incoming stream |
 | Client to server | `input_audio_buffer.append` | Appends base64 PCM16 audio |
 | Client to server | `input_audio_buffer.commit` with `final: true` | Marks the end of input |
 | Server to client | `transcription.delta` | Carries incremental response text |
 | Server to client | `transcription.done` | Carries final text and token usage |
 | Server to client | `response.audio.delta` | Carries incremental PCM16 response audio |
 | Server to client | `response.audio.done` | Marks the end of response audio |
 | Server to client | `error` | Reports an invalid event, model, or audio payload |
 A minimal client sends events in this order:
 ```json
 {"type":"session.update","model":"Qwen/Qwen3-Omni-30B-A3B-Instruct"}
 {"type":"input_audio_buffer.commit","final":false}
 {"type":"input_audio_buffer.append","audio":"<base64-pcm16>"}
 {"type":"input_audio_buffer.commit","final":true}
 ```
 The initial non-final commit intentionally starts generation before all audio
 has arrived. Continue sending `input_audio_buffer.append` events while the
 engine is consuming the stream.
 ## Audio Handling
 - Input is mono PCM16 at 16 kHz for the Qwen3-Omni example.
 - `response.audio.delta.audio` contains base64-encoded PCM16 bytes.
 - Read `sample_rate_hz` from each audio event instead of assuming an output
  rate. Qwen3-Omni output is typically 24 kHz.
 - Concatenate audio deltas in receive order to construct the output waveform.
 - Set `input_audio_buffer.commit.final` to `true` after the last input chunk so
  the server can terminate the streaming request.
 ## Availability and Limitations
 The path is registered on the unified API server, but it is usable only when
 the loaded pipeline implements realtime audio input and produces compatible
 audio output. Unsupported deployments return an `error` event. The endpoint
 does not provide duplex session resume, playback acknowledgement, overlap
 policy, or barge-in controls.
 See the [Qwen3-Omni online serving example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/qwen3_omni)
 for concurrency options, chunk pacing, and per-delta audio debugging.
@@ -0,0 +1,109 @@
 # Standalone Experimental Servers
 Most vLLM-Omni endpoints run on the unified server started with
 `vllm serve <model> --omni`. The experimental full-duplex package also contains
 two separate user-facing server processes. Their routes are not added to the
 unified API server.
 | Server | Transport | Role |
 |--------|-----------|------|
 | PersonaPlex compatibility server | WebSocket | Hosts the official PersonaPlex browser protocol and a raw-PCM alternative |
 | JoyVL interaction server | HTTP | Adds state, memory, proactive decisions, and delegation in front of a separate OpenAI-compatible model backend |
 ## PersonaPlex Compatibility Server
 Use this server when a client requires the official Moshi/PersonaPlex wire
 protocol or the bundled browser UI. For the engine-owned unified duplex path,
 use [`WS /v1/realtime?duplex=1`](full_duplex_api.md) instead.
 Start it with:
 ```bash
 HF_TOKEN=... CUDA_VISIBLE_DEVICES=0 python -m \
  vllm_omni.experimental.fullduplex.personaplex.serving.server \
  --port 8091
 ```
 | Route | Purpose |
 |-------|---------|
 | `GET /` | Official PersonaPlex browser client, when its assets are available |
 | `GET /health` | Readiness check |
 | `WS /api/chat` | Moshi-compatible tagged binary protocol with Opus audio |
 | `WS /v1/audio/duplex` | Simpler float32 raw-PCM protocol |
 ### `/api/chat` Protocol
 The first byte of every binary frame is a message tag:
 | Direction | Tag | Payload |
 |-----------|-----|---------|
 | Server to client | `0x00` | Ready handshake |
 | Client to server | `0x01` | Opus microphone audio at 24 kHz |
 | Server to client | `0x01` | Opus response audio |
 | Server to client | `0x02` | UTF-8 inner-monologue text |
 The optional `text_prompt` and `voice_prompt` query parameters select a persona
 and voice. The default server admits one conversation; use `--batch-size N` to
 configure multiple lockstep slots. Connections beyond capacity close with
 WebSocket code 1013.
 ### `/v1/audio/duplex` Protocol
 Send `{"type":"open","persona":"...","voice":"..."}` as a JSON text frame,
 then exchange binary float32 PCM frames. The server returns generated float32
 audio as binary frames and text as `{"type":"text","text":"..."}`. Send
 `{"type":"close"}` and wait for `{"type":"done"}` to flush and finish.
 See the [PersonaPlex example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/personaplex)
 for the browser and headless clients.
 ## JoyVL Interaction Server
 JoyVL is an orchestration layer, not a model-serving engine. It calls a
 separate OpenAI-compatible backend and adds per-session frame history, memory,
 persona policy, and optional delegation.
 Start the model backend and orchestrator separately:
 ```bash
 vllm serve jdopensource/JoyAI-VL-Interaction-Preview \
  --served-model-name JoyAI-VL-Interaction-Preview \
  --port 8092 \
  --max-model-len 131072 \
  --enable-prefix-caching \
  --limit-mm-per-prompt '{"image":256,"video":1}'
 python -m vllm_omni.experimental.fullduplex.joyvl.serving.server \
  --port 8091 \
  --main-backend-url http://127.0.0.1:8092/v1 \
  --main-model JoyAI-VL-Interaction-Preview
 ```
 | Method and route | Purpose |
 |------------------|---------|
 | `GET /health` | Readiness check |
 | `GET /v1/models` | Reports the configured interaction model |
 | `POST /v1/chat/completions` | Processes one frame or interaction tick |
 | `POST /reset` | Resets a session |
 | `POST /v1/streaming/reset` | Alias for `/reset` |
 | `POST /v1/streaming/persona` | Changes the session persona |
 Send one multimodal Chat Completions request per video frame, normally around
 one frame per second, and identify the session with `x-session-id`. Responses
 include an `interaction` block whose action is `silence`, `response`, or
 `delegate`.
 ```bash
 curl http://127.0.0.1:8091/v1/chat/completions \
  -H 'content-type: application/json' \
  -H 'x-session-id: demo' \
  -d '{
    "messages": [{"role": "user", "content": [
      {"type": "text", "text": "Alert me if a fire appears"},
      {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}
    ]}]
  }'
 ```
 See the [JoyVL recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/JD/JoyAI-VL-Interaction.md)
 for memory, personas, delegation backends, and the browser UI.
@@ -0,0 +1,92 @@
 # Streaming Video Output WebSocket API
 Use `WS /v1/realtime/video` to receive fragmented MP4 chunks while a compatible
 diffusion video model is generating. This endpoint streams generated video
 output; it is unrelated to [`WS /v1/video/chat/stream`](video_stream_api.md),
 which streams live video input into a conversational model.
 ## Quick Start
 Start a model that supports step execution and enable diffusion streaming
 output:
 ```bash
 vllm serve BestWishYsh/Helios-Distilled --omni \
  --diffusion-streaming-output \
  --port 8091
 ```
 Install the client dependency and run the provided client:
 ```bash
 pip install av websockets
 python examples/online_serving/streaming_video_generation/streaming_video_client.py \
  --host 127.0.0.1 \
  --port 8091 \
  --model BestWishYsh/Helios-Distilled \
  --prompt "A serene lakeside sunrise with mist over the water" \
  --output output.mp4
 ```
 ## Protocol
 Control events are JSON text frames. Each `video.chunk_metadata` event is
 immediately followed by one binary WebSocket frame containing fragmented MP4
 bytes.
 | Direction | Event or frame | Purpose |
 |-----------|----------------|---------|
 | Client to server | `session.start` | Starts generation with a prompt and video parameters |
 | Server to client | `video.start` | Confirms the request ID, format, and accepted configuration |
 | Server to client | `video.chunk_metadata` | Describes the next binary media frame |
 | Server to client | Binary frame | Carries one fragmented MP4 chunk |
 | Client to server | `session.interaction` | Queues an optional prompt update during generation |
 | Server to client | `session.interaction.queued` | Confirms that an interaction was queued |
 | Client to server | `session.ping` | Refreshes the stall clock |
 | Server to client | `session.pong` | Acknowledges a ping |
 | Client to server | `session.stop` | Cancels the active generation |
 | Server to client | `session.done` | Reports completion or cancellation and the chunk count |
 | Server to client | `error` | Reports validation, generation, or timeout failures |
 Start a minimal session with:
 ```json
 {
  "type": "session.start",
  "model": "BestWishYsh/Helios-Distilled",
  "prompt": "A serene lakeside sunrise",
  "format": "m4s",
  "width": 640,
  "height": 384,
  "fps": 16,
  "num_frames": 99
 }
 ```
 Only `m4s` output is currently accepted. Concatenate the binary frames in
 receive order, then remux the fragmented stream if the target player requires
 a progressive MP4 file.
 ## Mid-Generation Interaction
 Models that implement interaction updates can change the active prompt:
 ```json
 {
  "type": "session.interaction",
  "interaction": {
    "event_id": "sunlight-update",
    "event": {"prompt": "Sunlight breaks through the mist"},
    "transition_chunks": 2
  }
 }
 ```
 This feature is pipeline-dependent. An enabled endpoint does not imply that
 the loaded model supports midway prompt updates.
 The default start timeout is 10 seconds and the server reports a stall after
 about 60 seconds without generation progress or a `session.ping`. See the
 [streaming video generation example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/streaming_video_generation)
 for prompt-update scheduling and browser playback.
@@ -17,6 +17,8 @@ globally.
 | --- | --- | --- |
 | `sparsity` | finite, `[0, 1]` | Nominal dropped-key-block fraction; default `0.8`; `0` disables sparsity |
 | `start_step` | integer, `>= 0` | Number of early denoise steps kept dense |
 | `end_step` | integer, `>= 0` | Number of final denoise steps kept dense (tail fallback) |
 | `precision` | `"bf16"`, `"fp8"`, `"mix"` | Kernel precision mode; default `"bf16"`. Requires MindIE-SD with `sparse_attention(precision=...)` |
 | `skip_layers` | selector such as `"0-3,38"` | DiT blocks kept dense |
 ```bash
@@ -52,6 +54,12 @@ platform raises. It is incompatible with ring sequence parallelism because
 the kernel needs the complete key sequence for ranking; use Ulysses sequence
 parallelism with `ring_degree=1`.
 The `precision` knob requires a MindIE-SD release whose `sparse_attention`
 accepts `precision=`; older releases accept it through `**kwargs` but
 silently ignore it and run the BF16 path. `RAINFUSION_ATTN` raises at runtime
 when a non-`bf16` precision is requested without that support, so pin a
 compatible MindIE-SD release when using `precision="mix"` or `"fp8"`.
 ## Geometry handling
 RainFusion handles arbitrary video grids by rearranging video spatially and,
@@ -12,15 +12,17 @@ LoRA adapters are lightweight, model-specific fine-tuning weights that can be ap
 ## LoRA Adapter Format
 ### PEFT (Parameter-Efficient Fine-Tuning) format (default)
 A typical PEFT-format LoRA adapter directory structure:
-```
+```text
 lora_adapter/
 ├── adapter_config.json
 └── adapter_model.safetensors
 ```
 The `adapter_config.json` file contains metadata about the LoRA adapter, including:
 - `r`: LoRA rank
 - `lora_alpha`: LoRA alpha scaling factor
 - `target_modules`: List of module names to apply LoRA to
@@ -67,7 +69,7 @@ For distilled few-step LoRAs, pass `lora_backend="distill"` together with one or
 For Qwen-Image and Wan pipelines, the distill backend calls `pipeline.load_lora_weights(...)` during worker initialization.
 | Pipeline | Supported distilled LoRA repo | Notes |
-|----------|--------------------------------|-------|
+| ---------- | -------------------------------- | ------- |
 | `QwenImagePipeline` | `lightx2v/Qwen-Image-2512-Lightning` | Used with Qwen-Image-2512 Lightning-style few-step inference. |
 | `Wan22Pipeline` | `lightx2v/Wan2.1-Distill-Loras`, `lightx2v/Wan2.2-Distill-Loras` | Wan2.1 uses one LoRA file. For dual-transformer Wan2.2 MoE, pass high-noise then low-noise LoRA files. |
 | `Wan22I2VPipeline` | `lightx2v/Wan2.2-Distill-Loras` | For dual-transformer Wan2.2 MoE, pass high-noise then low-noise LoRA files. |
@@ -240,6 +242,37 @@ Notes:
 - Output quality and speed depend on the replacement checkpoints and sampling params you choose.
 - If you only need to fuse distilled LoRAs into a Wan2.2 checkpoint at load time (without the full LightX2V convert + assemble pipeline), you can instead pass them directly via `--lora-backend distill --lora-path <high>.safetensors <low>.safetensors`. See the [Distill backend](#distill-backend-fuse-distilled-lora-at-init) section above.
 ## MiniMax-H3 adapter-declared schedules
 MiniMax-H3 supports two few-step mechanisms that must not be conflated:
 - **Checkpoint-pinned schedule**: a merged release writes `base_schedule` into
  `model_index.json`. Requests must pass `num_inference_steps` as the interval
  count (for example `4` for `[1.0, 0.7, 0.4, 0.15, 0.0]`).
 - **Runtime Turbo LoRA**: LightX2V Turbo artifacts request five sigma points and
  enforce `flow_shift=6`, `audio_flow_shift=3`.
 - **Runtime native LoRA**: FlashGen-style artifacts declare
  `key_format=minimax-h3-native` and embed `base_schedule` in safetensors
  metadata. When active, the adapter schedule overrides the base checkpoint and
  requests must use the interval-count contract (`num_inference_steps=4`).
  Request-mode generation may omit the field and take the count from the adapter
  schedule; step execution requires it explicitly, because the step scheduler
  derives the total step count from the request before the adapter schedule is
  known.
 Native artifacts also declare `qkv_layout=grouped`. The H3 loader reorders fused
 `qkv_proj` LoRA rows with the same `_reorder_grouped_qkv_to_qkv` helper used for
 base-weight loading, then binds the packed Q/K/V slices through the legacy PEFT
 manager without modifying `DiffusionLoRAManager`.
 Both runtime adapters run with distributed layerwise offload in request-mode
 generation, where the manager keeps the LoRA A/B buffers resident on the compute
 device while DLO streams the base blocks. MiniMax-H3 step execution still
 rejects DLO, so `--step-execution` cannot be combined with
 `--enable-distributed-layerwise-offload`. Model-level CPU offload and standard
 layerwise offload are rejected in every mode, because the dynamic LoRA tensors
 are neither parameters nor registered buffers and therefore do not participate
 in those weight lifecycles.
 ## See Also
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 Image-to-Video generation example using Wan2.2 I2V/TI2V models, LTX2/LTX-2.3,
@@ -111,6 +111,13 @@ def build_image_to_video_prompt(
    return result
 def _validate_video_output_type(output_type: str) -> None:
    if output_type not in {"image", "video"}:
        raise ValueError(
            f"Unexpected output type '{output_type}', expected 'video' or legacy 'image' for video generation."
        )
 def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description=(
@@ -665,10 +672,7 @@ def main():
        frames = frames[0] if frames else None
    if isinstance(frames, OmniRequestOutput):
-        if frames.final_output_type != "image":
+        _validate_video_output_type(frames.final_output_type)
            raise ValueError(
                f"Unexpected output type '{frames.final_output_type}', expected 'image' for video generation."
            )
        if frames.multimodal_output and "audio" in frames.multimodal_output:
            audio = frames.multimodal_output["audio"]
            audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 import argparse
 import json
@@ -174,6 +174,13 @@ def build_text_to_video_prompt(prompt: str, negative_prompt: str | None) -> dict
    return result
 def _validate_video_output_type(output_type: str) -> None:
    if output_type not in {"image", "video"}:
        raise ValueError(
            f"Unexpected output type '{output_type}', expected 'video' or legacy 'image' for video generation."
        )
 def _normalize_float_tensor(tensor: torch.Tensor, source_range: str) -> torch.Tensor:
    """Normalize decoded frames according to the pipeline's output contract."""
    if not tensor.is_floating_point():
@@ -620,10 +627,7 @@ def main():
        frames = frames[0] if frames else None
    if isinstance(frames, OmniRequestOutput):
-        if frames.final_output_type != "image":
+        _validate_video_output_type(frames.final_output_type)
            raise ValueError(
                f"Unexpected output type '{frames.final_output_type}', expected 'image' for video generation."
            )
        if frames.multimodal_output and "audio" in frames.multimodal_output:
            audio = frames.multimodal_output["audio"]
            audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
@@ -49,7 +49,7 @@ scenario below for live barge-in validation on the target GPU.
 Default:
 ```bash
-vllm-omni serve openbmb/MiniCPM-o-4_5 \
+vllm serve openbmb/MiniCPM-o-4_5 \
    --omni \
    --deploy-config vllm_omni/deploy/minicpmo_4_5.yaml \
    --trust-remote-code \
@@ -159,6 +159,33 @@ python examples/online_serving/minicpmo/realtime_duplex_demo.py \
    --output-dir /tmp/minicpmo_realtime_duplex_demo
 ```
 Video input uses the same session. PyAV demuxes JPEG frames at `--video-fps`
 (default 1.0) and vLLM `load_audio` extracts a 16 kHz mono WAV unless
 `--input-wav` overrides the soundtrack. Frames keep their capture size
 (`--frame-max-side 0`); the server `process_image` normalizes at 448.
 `--stack-frames N` raises the visual refresh rate the way official duplex does:
 each 1 s unit also samples the `N-1` sub-frames captured inside it and tiles
 them into a single composite image sent next to that unit's base frame. The
 audio cadence never changes — a unit is always one second — and the wire always
 carries 2 images per unit however large `N` is, because the sub-frames share one
 composite. Official uses 5 for high refresh rate mode:
 ```bash
 python examples/online_serving/minicpmo/realtime_duplex_demo.py \
    --url ws://localhost:8099/v1/realtime?duplex=1 \
    --model openbmb/MiniCPM-o-4_5 \
    --input-video /path/to/clip.mp4 \
    --stack-frames 5 \
    --ref-audio /path/to/MiniCPM-o-Demo/assets/ref_audio/ref_minicpm_signature.wav \
    --output-dir /tmp/minicpmo_realtime_duplex_video_demo
 ```
 Detail inside a composite is capped by `scale_resolution=448` at
 `max_slice_nums=1`: official suggests HD slicing (`max_slice_nums=[2, 1]`) for
 stacked frames, which the duplex adapter does not implement yet. Reading small
 text or digits out of a wide scene is limited by that, not by frame timing.
 ## Open the experimental browser client
 The browser UI serves the page and proxies the same-origin Realtime WebSocket to
@@ -12,7 +12,9 @@ from __future__ import annotations
 import argparse
 import asyncio
 import base64
 import io
 import json
 import math
 import sys
 import time
 from pathlib import Path
@@ -22,8 +24,8 @@ if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))
 from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
-    PCM16_BYTES_PER_SAMPLE,  # noqa: F401 - compatibility export
+    PCM16_BYTES_PER_SAMPLE,
-    PCM16_SAMPLE_RATE,  # noqa: F401 - compatibility export
+    PCM16_SAMPLE_RATE,
    RealtimeDuplexClient,
    RealtimeEventCollector,
    build_realtime_url,
@@ -38,6 +40,10 @@ from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
 from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
    reference_audio_data_url as _ref_audio_data_url,
 )
 from vllm_omni.experimental.fullduplex.video_stacking import (  # noqa: E402
    concat_frames_b64,
    unit_subframe_offsets,
 )
 class _StreamingOutputWriter:
@@ -149,6 +155,23 @@ def _response_in_progress(events: list[dict[str, object]]) -> bool:
    )
 async def _commit_input_after_playback_checkpoint(client: RealtimeDuplexClient) -> float:
    """Commit input only after preserving already-played assistant history.
    Native duplex can finish several responses while the user is still
    streaming.  Under the ``ack_only`` playback policy, acknowledging those
    responses after the final input commit is intentionally rejected: it could
    append an old assistant message after the newly committed user message.
    Checkpoint playback first so the server commits each assistant item at its
    original history position.  A second acknowledgement after the commit can
    then safely advance response-local playback that drained in the meantime.
    """
    await client.acknowledge_playback()
    commit_sent_at_s = time.monotonic()
    await client.commit()
    return commit_sent_at_s
 def _event_count_after(
    events: list[dict[str, object]],
    event_type: str,
@@ -159,14 +182,258 @@ def _event_count_after(
    return sum(event.get("type") == event_type for event in events[index + 1 :])
-async def run_demo(args: argparse.Namespace) -> dict[str, object]:
+def _probe_video_frames_and_fps(video_path: Path) -> tuple[int, float]:
-    input_pcm16 = read_pcm16_wav(Path(args.input_wav))
+    """Frame count and average FPS via PyAV (same approach as Daily-Omni)."""
-    if not input_pcm16:
+    import av
        raise ValueError("input WAV has no audio")
    with av.open(str(video_path)) as container:
        stream = container.streams.video[0]
        avg_fps = float(stream.average_rate) if stream.average_rate else 0.0
        num_frames = stream.frames or 0
        if num_frames <= 0:
            duration: float | None = None
            if stream.duration is not None and stream.time_base is not None:
                duration = float(stream.duration * stream.time_base)
            elif container.duration is not None:
                duration = container.duration / av.time_base
            if duration and avg_fps:
                num_frames = int(math.floor(duration * avg_fps))
        if num_frames <= 0:
            stream.thread_type = "AUTO"
            num_frames = sum(1 for _ in container.decode(stream))
    if num_frames <= 0:
        raise ValueError(f"No decodable video frames in {video_path}")
    return num_frames, avg_fps or 1.0
 def _upright(image, rotation: int):
    """Apply a container display-matrix rotation, in degrees, to ``image``.
    PyAV hands back the stored raster and leaves the display matrix for the
    caller, unlike the ffmpeg CLI which autorotates. Phone clips routinely
    carry +/-90, so skipping this feeds the model a sideways frame: arrows and
    digits stop being readable and the model mirrors the motion it reports.
    """
    from PIL import Image
    quarter_turns = {
        90: Image.Transpose.ROTATE_90,
        180: Image.Transpose.ROTATE_180,
        270: Image.Transpose.ROTATE_270,
    }
    turn = quarter_turns.get(int(rotation) % 360)
    return image.transpose(turn) if turn is not None else image
 def _decode_video_frames_rgb(video_path: Path, frame_idx: list[int]) -> list:
    """Decode presentation-order ``frame_idx`` into upright RGB ``PIL.Image``s."""
    import av
    wanted = sorted({int(i) for i in frame_idx})
    if not wanted:
        return []
    decoded: dict[int, object] = {}
    with av.open(str(video_path)) as container:
        stream = container.streams.video[0]
        stream.thread_type = "AUTO"
        cursor = 0
        for position, frame in enumerate(container.decode(stream)):
            if wanted[cursor] > position:
                continue
            image = _upright(frame.to_image(), getattr(frame, "rotation", 0) or 0)
            while cursor < len(wanted) and wanted[cursor] <= position:
                decoded[wanted[cursor]] = image
                cursor += 1
            if cursor >= len(wanted):
                break
    if not decoded:
        raise ValueError(f"Decoded no frames from {video_path} for indices {wanted[:8]}")
    last_image = decoded[max(decoded)]
    return [decoded.get(int(i), last_image) for i in frame_idx]
 def _resize_max_side(image, max_side: int):
    width, height = image.size
    longest = max(width, height)
    if longest <= max_side or max_side <= 0:
        return image
    scale = max_side / float(longest)
    new_size = (max(1, int(round(width * scale))), max(1, int(round(height * scale))))
    return image.resize(new_size)
 def _jpeg_bytes(image, *, quality: int = 95) -> bytes:
    buf = io.BytesIO()
    image.convert("RGB").save(buf, format="JPEG", quality=quality)
    return buf.getvalue()
 def _to_pcm16_bytes(audio_np) -> bytes:
    import numpy as np
    arr = np.asarray(audio_np).reshape(-1)
    if arr.dtype != np.int16:
        arr = (np.clip(arr, -1.0, 1.0) * 32767.0).astype(np.int16)
    return arr.tobytes()
 def _pcm16_wav_duration_s(path: Path) -> float:
    pcm16 = read_pcm16_wav(path)
    return len(pcm16) / float(PCM16_SAMPLE_RATE * PCM16_BYTES_PER_SAMPLE)
 def _extract_video_frames(
    video_path: Path,
    work_dir: Path,
    *,
    fps: float,
    max_side: int,
    duration_s: float,
    stack_frames: int = 1,
 ) -> tuple[list[str], list[str | None]]:
    """Demux JPEG frames only. Frame count follows ``duration_s``, not soundtrack.
    Returns the base frames (one per model unit) and the parallel stacked track.
    With ``stack_frames=N`` each unit also samples its ``N-1`` interior
    sub-frames and tiles them into one composite, which is how official duplex
    raises the visual refresh rate while keeping one second of audio per unit.
    A unit whose interior falls past the end of the clip stacks nothing.
    """
    num_frames, avg_fps = _probe_video_frames_and_fps(video_path)
    sample_fps = max(float(fps), 1e-6)
    unit_s = 1.0 / sample_fps
    timeline_s = max(float(duration_s), unit_s)
    num_samples = max(1, int(math.ceil(timeline_s * sample_fps - 1e-9)))
    offsets = unit_subframe_offsets(stack_frames, unit_s=unit_s)
    def frame_at(seconds: float) -> int:
        # Official get_video_frame_audio_segments uses truncating int(), not
        # round(), so 29.97 fps lands on the same presentation index.
        return min(int(seconds * avg_fps), num_frames - 1)
    base_idx = [frame_at(unit * unit_s) for unit in range(num_samples)]
    sub_times = [
        [unit * unit_s + offset for offset in offsets if unit * unit_s + offset < timeline_s]
        for unit in range(num_samples)
    ]
    sub_idx = [frame_at(seconds) for times in sub_times for seconds in times]
    images = _decode_video_frames_rgb(video_path, base_idx + sub_idx)
    sub_images = iter(images[len(base_idx) :])
    frame_dir = work_dir / "frames"
    frame_dir.mkdir(parents=True, exist_ok=True)
    for stale in frame_dir.glob("*.jpg"):
        stale.unlink()
    frames_b64: list[str] = []
    stacked_b64: list[str | None] = []
    for index, (image, times) in enumerate(zip(images[: len(base_idx)], sub_times), start=1):
        jpeg = _jpeg_bytes(_resize_max_side(image, max_side))
        (frame_dir / f"frame_{index:05d}.jpg").write_bytes(jpeg)
        frames_b64.append(base64.b64encode(jpeg).decode("ascii"))
        interior = [_resize_max_side(next(sub_images), max_side) for _ in times]
        if not interior:
            stacked_b64.append(None)
            continue
        composite = concat_frames_b64(interior)
        (frame_dir / f"stack_{index:05d}.jpg").write_bytes(base64.b64decode(composite))
        stacked_b64.append(composite)
    return frames_b64, stacked_b64
 def _extract_video_soundtrack(video_path: Path, work_dir: Path) -> Path:
    """Write a 16 kHz mono WAV from the video soundtrack."""
    try:
        from vllm.multimodal.media.audio import load_audio
    except ImportError as exc:  # pragma: no cover
        raise SystemExit("--input-video needs vLLM audio helpers (load_audio)") from exc
    audio_np, sr = load_audio(str(video_path), sr=PCM16_SAMPLE_RATE, mono=True)
    work_dir.mkdir(parents=True, exist_ok=True)
    audio_path = work_dir / "input_16k.wav"
    write_pcm16_wav(audio_path, _to_pcm16_bytes(audio_np), sample_rate_hz=int(sr))
    return audio_path
 def _extract_video_input(
    video_path: Path,
    work_dir: Path,
    *,
    fps: float,
    max_side: int,
    stack_frames: int,
 ) -> tuple[Path, list[str], list[str | None]]:
    """Demux a video into duplex inputs: 16 kHz mono WAV plus JPEG frames.
    Use this only when the video soundtrack is the audio source. When the
    caller already has ``--input-wav``, extract frames against that WAV
    duration instead so a silent clip cannot fail ``load_audio``.
    """
    audio_path = _extract_video_soundtrack(video_path, work_dir)
    frames_b64, stacked_b64 = _extract_video_frames(
        video_path,
        work_dir,
        fps=fps,
        max_side=max_side,
        duration_s=_pcm16_wav_duration_s(audio_path),
        stack_frames=stack_frames,
    )
    return audio_path, frames_b64, stacked_b64
 def _resolve_duplex_av_inputs(
    *,
    input_wav: str | None,
    input_video: str | None,
    work_dir: Path,
    fps: float,
    max_side: int,
    stack_frames: int = 1,
 ) -> tuple[str, list[str], list[str | None]]:
    """Return ``(wav_path, frames_b64, stacked_b64)``. External WAV wins over video audio."""
    if input_video:
        video_path = Path(input_video)
        if input_wav:
            frames, stacked = _extract_video_frames(
                video_path,
                work_dir,
                fps=fps,
                max_side=max_side,
                duration_s=_pcm16_wav_duration_s(Path(input_wav)),
                stack_frames=stack_frames,
            )
            return input_wav, frames, stacked
        extracted_wav, frames, stacked = _extract_video_input(
            video_path,
            work_dir,
            fps=fps,
            max_side=max_side,
            stack_frames=stack_frames,
        )
        return str(extracted_wav), frames, stacked
    if not input_wav:
        raise SystemExit("provide --input-wav, --input-video, or both")
    return input_wav, [], []
 async def run_demo(args: argparse.Namespace) -> dict[str, object]:
    output_dir = Path(args.output_dir)
    stream_writer = _StreamingOutputWriter(output_dir)
    input_wav, video_frames, stacked_video_frames = _resolve_duplex_av_inputs(
        input_wav=args.input_wav,
        input_video=args.input_video,
        work_dir=output_dir / "video_input",
        fps=args.video_fps,
        max_side=args.frame_max_side,
        stack_frames=args.stack_frames,
    )
    input_pcm16 = read_pcm16_wav(Path(input_wav))
    if not input_pcm16:
        raise ValueError("input WAV has no audio")
    url = build_realtime_url(
        args.url,
        args.model,
@@ -184,10 +451,12 @@ async def run_demo(args: argparse.Namespace) -> dict[str, object]:
            timeout_s=args.timeout_s,
        )
        stream_event_cursor = len(client.events.events)
-        await client.stream_pcm16(
+        frames_sent = await client.stream_pcm16(
            input_pcm16,
            chunk_ms=args.chunk_ms,
            realtime=not args.no_realtime_pacing,
            video_frames=video_frames,
            stacked_video_frames=stacked_video_frames,
        )
        commit_event_cursor = len(client.events.events)
        stream_decision = _latest_model_decision(client.events.events, stream_event_cursor)
@@ -196,8 +465,7 @@ async def run_demo(args: argparse.Namespace) -> dict[str, object]:
            chunk_period_ms=_chunk_period_ms(client.events.events),
        )
        wait_for_post_commit_decision = False
-        commit_sent_at_s = time.monotonic()
+        commit_sent_at_s = await _commit_input_after_playback_checkpoint(client)
        await client.commit()
        wait_error: str | None = None
        committed_index: int | None = None
        post_commit_decision: str | None = None
@@ -310,6 +578,10 @@ async def run_demo(args: argparse.Namespace) -> dict[str, object]:
                    committed_index,
                ),
            },
            "input_video": args.input_video,
            "video_frames_extracted": len(video_frames),
            "video_frames_sent": frames_sent,
            "video_stacked_frames_extracted": sum(frame is not None for frame in stacked_video_frames),
            "audio_bytes": len(audio),
            "audio_chunk_count": len(stream_writer.audio_chunk_paths),
            "audio_chunk_files": [str(path) for path in stream_writer.audio_chunk_paths],
@@ -356,7 +628,48 @@ def parse_args() -> argparse.Namespace:
    parser.add_argument("--url", default="ws://localhost:8099/v1/realtime?duplex=1")
    parser.add_argument("--model", default="openbmb/MiniCPM-o-4_5")
    parser.add_argument("--session-id")
-    parser.add_argument("--input-wav", required=True)
+    parser.add_argument(
        "--input-wav",
        help="Mono 16 kHz PCM16 clip to stream. Optional when --input-video carries the audio track.",
    )
    parser.add_argument(
        "--input-video",
        help=(
            "Video to stream as omni duplex input. PyAV demuxes frames and "
            "vLLM load_audio extracts a 16 kHz mono WAV (unless --input-wav overrides the audio)."
        ),
    )
    parser.add_argument(
        "--video-fps",
        type=float,
        default=1.0,
        help="Frame extraction rate. The wire contract carries ~1 frame per second of audio.",
    )
    parser.add_argument(
        "--frame-max-side",
        type=int,
        default=0,
        help=(
            "Optional client-side downscale of the longest side, in pixels. "
            "0 (default) keeps the capture resolution so the server "
            "process_image can normalize at scale_resolution=448 from a "
            "higher-quality source. The web demo never resizes on the client."
        ),
    )
    parser.add_argument(
        "--stack-frames",
        type=int,
        default=1,
        help=(
            "Frames sampled per 1 s audio unit. 1 (default) sends the unit's "
            "frame alone. N > 1 also samples the N-1 sub-frames inside the unit "
            "and tiles them into one composite sent next to the base frame, so "
            "motion (arrows, doors, digits changing) is visible at N fps while "
            "the audio stays 1 s per unit. Official uses 5 for high refresh "
            "rate mode; the wire carries 2 images per unit at any N because the "
            "sub-frames share one composite."
        ),
    )
    parser.add_argument(
        "--ref-audio",
        required=True,
@@ -372,7 +685,7 @@ def parse_args() -> argparse.Namespace:
        "--temperature",
        type=float,
        default=None,
-        help="Stage0 sampling temperature; response-required soft-interrupt uses 0.0.",
+        help="Stage 0 sampling temperature; omit to preserve the model default.",
    )
    parser.add_argument("--no-realtime-pacing", action="store_true")
    parser.add_argument("--require-audio", action="store_true")
@@ -37,7 +37,7 @@ During generation the client normally sends only `session.start` and then receiv
 ## Install Client Dependency
 ```bash
-pip install websockets
+pip install av websockets
 # For the Gradio demo:
 pip install 'vllm-omni[demo]' websockets
 ```
@@ -8,7 +8,7 @@ the received bytes to disk when the session finishes or is interrupted.
 Image/reference input is intentionally not supported in this example yet.
 Requirements:
-    pip install websockets
+    pip install av websockets
 """
 import argparse
@@ -71,11 +71,11 @@ dev = [
    "voxcpm==2.0.3", # for VoxCPM2 path-locating in tests
    "torchdiffeq==0.2.5", # for covo_audio Test
    "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)
    "silero-vad==6.1", # Realtime server-VAD duplex E2E
 ]
 demo = [
@@ -85,6 +85,10 @@ demo = [
    "cn2an>=0.5.22",  # Chinese number normalization for TTS (e.g. "1次" -> "一次")
 ]
 server-vad = [
    "silero-vad==6.1",
 ]
 forced-aligner = [
    "qwen-asr>=0.0.6",  # faithful Japanese/Korean word segmentation for the TTS forced aligner
 ]
@@ -244,11 +248,11 @@ addopts = [
    "--dist=loadgroup"
 ]
 markers = [
-    # ci/cd required
+    # ci/cd required — tag with [ci-level] for tests.helpers.mark.get_level_markers().
-    "core_model: L1&L2 tests (run in each PR)",
+    "core_model: [ci-level] L1&L2 tests (run in each PR)",
-    "advanced_model: L3 level tests (run on each merge)",
+    "advanced_model: [ci-level] L3 level tests (run on each merge)",
-    "full_model: L4 level tests (run nightly)",
+    "full_model: [ci-level] L4 level tests (run nightly)",
-    "local_model: tests (run locally)",
+    "local_model: [ci-level] tests (run locally)",
    # function module markers
    "diffusion: Diffusion model tests",
    "omni: Omni model tests",
@@ -258,38 +262,41 @@ markers = [
    "sp: Sequence parallelism tests",
    "example: Doc example code tests",
    # platform markers
-    "cpu: Tests that run on CPU",
+    # Tag with [hardware-platform] for tests.helpers.mark.get_supported_platforms().
-    "gpu: Tests that run on GPU (auto-added)",
+    # Add [gpu] on accelerator platforms that should also get pytest.mark.gpu.
-    "cuda: Tests that run on CUDA (auto-added)",
+    # cpu/gpu are pytest marks; hardware_test(res=...) rejects them as keys.
-    "rocm: Tests that run on AMD/ROCm (auto-added)",
+    "cpu: [hardware-platform] Tests that run on CPU",
-    "xpu: Tests that run on XPU (auto-added)",
+    "gpu: [hardware-platform] Tests that run on GPU (auto-added)",
-    "npu: Tests that run on NPU/Ascend (auto-added)",
+    "cuda: [hardware-platform] [gpu] Tests that run on CUDA (auto-added)",
-    "musa: Tests that run on MUSA/Moore Threads (auto-added)",
+    "rocm: [hardware-platform] [gpu] Tests that run on AMD/ROCm (auto-added)",
    "xpu: [hardware-platform] [gpu] Tests that run on XPU (auto-added)",
    "npu: [hardware-platform] Tests that run on NPU/Ascend (auto-added)",
    "musa: [hardware-platform] [gpu] Tests that run on MUSA/Moore Threads (auto-added)",
    # specified computation resources marks (auto-added)
-    # Tag with [hardware-resource] for tests.helpers.mark.get_hardware_mark_list().
+    # [hardware-resource] + [cuda]/[rocm]/[xpu]/[npu]/[musa] for hardware_test(res=...).
-    "H100: [hardware-resource] Tests that require H100 GPU",
+    # Do not put [gpu] on SKU lines: that would make gpu a res key.
-    "H800: [hardware-resource] Tests that require H800 GPU",
+    "H100: [hardware-resource] [cuda] Tests that require H100 GPU",
-    "H200: [hardware-resource] Tests that require H200 GPU",
+    "H800: [hardware-resource] [cuda] Tests that require H800 GPU",
-    "L4: [hardware-resource] Tests that require L4 GPU",
+    "H200: [hardware-resource] [cuda] Tests that require H200 GPU",
-    "B200: [hardware-resource] Tests that require B200 GPU",
+    "L4: [hardware-resource] [cuda] Tests that require L4 GPU",
-    "MI325: [hardware-resource] Tests that require MI325 GPU (AMD/ROCm)",
+    "B200: [hardware-resource] [cuda] Tests that require B200 GPU",
-    "B60: [hardware-resource] Tests that require Intel Arc Pro B60 XPU",
+    "MI325: [hardware-resource] [rocm] Tests that require MI325 GPU (AMD/ROCm)",
-    "S5000: [hardware-resource] Tests that require S5000 GPU (Moore Threads/MUSA)",
+    "B60: [hardware-resource] [xpu] Tests that require Intel Arc Pro B60 XPU",
-    "A2: [hardware-resource] Tests that require A2 NPU",
+    "S5000: [hardware-resource] [musa] Tests that require S5000 GPU (Moore Threads/MUSA)",
-    "A3: [hardware-resource] Tests that require A3 NPU",
+    "A2: [hardware-resource] [npu] Tests that require A2 NPU",
-    "distributed_cuda: Tests that require multi cards on CUDA platform",
+    "A3: [hardware-resource] [npu] Tests that require A3 NPU",
-    "distributed_rocm: Tests that require multi cards on ROCm platform",
+    # Tag with [hardware-cards] for tests.helpers.mark.get_supported_card_counts().
-    "distributed_xpu: Tests that require multi cards on XPU platform",
+    "cards_1: [hardware-cards] Tests that require 1 accelerator card",
-    "distributed_npu: Tests that require multi cards on NPU platform",
+    "cards_2: [hardware-cards] Tests that require 2 accelerator cards",
-    "distributed_musa: Tests that require multi cards on MUSA platform",
+    "cards_3: [hardware-cards] Tests that require 3 accelerator cards",
-    "skipif_cuda: Skip if the num of CUDA cards is less than the required",
+    "cards_4: [hardware-cards] Tests that require 4 accelerator cards",
-    "skipif_rocm: Skip if the num of ROCm cards is less than the required",
+    "cards_5: [hardware-cards] Tests that require 5 accelerator cards",
-    "skipif_xpu: Skip if the num of XPU cards is less than the required",
+    "cards_6: [hardware-cards] Tests that require 6 accelerator cards",
-    "skipif_npu: Skip if the num of NPU cards is less than the required",
+    "cards_7: [hardware-cards] Tests that require 7 accelerator cards",
-    "skipif_musa: Skip if the num of MUSA cards is less than the required",
+    "cards_8: [hardware-cards] Tests that require 8 accelerator cards",
    # more detailed markers
    "sp: Sequence parallelism tests (multi-GPU)",
-    "slow: Slow tests (may skip in quick CI)",
+    "slow: [ci-level] Slow tests (may skip in quick CI)",
    "benchmark: Benchmark tests",
    "xdist: xdist-parallelizable tests",
 ]
@@ -1,11 +1,14 @@
 # Boogu-Image
-> Text-to-image and image-editing online serving (Boogu-Image-0.1-Base / -Edit)
+> Text-to-image and image-editing online serving
 (Boogu-Image-0.1-Base / -Edit / -Edit-Turbo)
 ## Summary
 - Vendor: Boogu
- Model: `Boogu/Boogu-Image-0.1-Base` (text-to-image), `Boogu/Boogu-Image-0.1-Edit` (image editing)
+- Model: `Boogu/Boogu-Image-0.1-Base` (text-to-image),
  `Boogu/Boogu-Image-0.1-Edit` (image editing), and
  `Boogu/Boogu-Image-0.1-Edit-Turbo` (four-step image editing)
 - Task: Text-to-image generation and text-guided image editing (TI2I)
 - Mode: Online serving with the OpenAI-compatible API
 - Maintainer: Community
@@ -13,19 +16,23 @@
 ## When to use this recipe
 Use this recipe when you want a known-good starting point for serving
-`Boogu/Boogu-Image-0.1-Base` with vLLM-Omni's native pipeline (no
+`Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit`, or
 `Boogu/Boogu-Image-0.1-Edit-Turbo` with vLLM-Omni's native pipeline (no
 `--diffusion-load-format diffusers`, and the upstream `boogu` package is not
 required).
 Boogu-Image-0.1 is an Apache-2.0 unified image generation and editing model
-family. This recipe covers the Base text-to-image checkpoint, which pairs a
+family. The Base text-to-image checkpoint pairs a Qwen3-VL multimodal encoder
-Qwen3-VL multimodal encoder with a Diffusion Transformer (DiT) and a flow-match
+with a Diffusion Transformer (DiT) and a flow-match Euler scheduler with
-Euler scheduler with time-shift. It handles photorealistic generation and
+time-shift. It handles photorealistic generation and Chinese/English text
-Chinese/English text rendering.
+rendering. The Edit and Edit-Turbo checkpoints use the same native
 `BooguImagePipeline` for image editing.
 ## References
 - Upstream model card: <https://huggingface.co/Boogu/Boogu-Image-0.1-Base>
 - Edit model card: <https://huggingface.co/Boogu/Boogu-Image-0.1-Edit>
 - Edit-Turbo 1K hotfix: <https://huggingface.co/Boogu/Boogu-Image-0.1-Edit-Turbo/tree/hotfix-1k-20260708>
 - Project page: <https://boogu.org>
 - GitHub: <https://github.com/boogu-project/Boogu-Image>
 - Related example: [`examples/online_serving/text_to_image/`](../../examples/online_serving/text_to_image/README.md)
@@ -93,7 +100,7 @@ curl -s http://localhost:8091/v1/chat/completions \
 - **Memory usage:** ~34.6 GiB on GPU. Use `--vae-use-slicing --vae-use-tiling`
  to trim peak VRAM if needed.
 - **Key flags:**
-  - `--omni` — enables vLLM-Omni diffusion serving.
+    - `--omni` — enables vLLM-Omni diffusion serving.
 - **Guidance:** Boogu-Image uses `guidance_scale` (mapped to the upstream
  `text_guidance_scale`); the default is `4.0`. Classifier-free guidance is
  active whenever `guidance_scale > 1.0`.
@@ -111,13 +118,13 @@ the image-editing (TI2I) path activates automatically when a request carries a
 reference image. The Base text-to-image path is unaffected (no reference image
 is sent).
-#### Command
+### Command
 ```bash
 vllm serve Boogu/Boogu-Image-0.1-Edit --omni --port 8091
 ```
-#### Verification
+### Verification
 Edit an image with `/v1/images/edits` (the model-card example — change a photo
 to a colored-pencil drawing). Diffusion parameters are plain multipart form
@@ -157,14 +164,14 @@ curl -s http://localhost:8091/v1/chat/completions \
  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2- | base64 -d > edited.png
 ```
-#### Notes
+### Notes
 - **Single reference image:** only one input image is supported for now (the
  upstream "Only support 1 reference image for now" limit).
 - **Guidance semantics:**
-  - `guidance_scale` = text guidance (upstream `text_guidance_scale`, default
+    - `guidance_scale` = text guidance (upstream `text_guidance_scale`, default
    `4.0`); `> 1.0` enables text CFG. Editing typically uses `5.0`.
-  - `guidance_scale_2` = image guidance (upstream `image_guidance_scale`,
+    - `guidance_scale_2` = image guidance (upstream `image_guidance_scale`,
    default `1.0` = off). Setting it `> 1.0` enables the double-guidance path
    (3 model predictions per step), steering more strongly toward the reference
    image.
@@ -178,3 +185,69 @@ curl -s http://localhost:8091/v1/chat/completions \
  supports Boogu-Image-Edit directly
  (`--model Boogu/Boogu-Image-0.1-Edit --guidance-scale 5.0`, optional
  `--guidance-scale-2`).
 ## Fast image editing (Boogu-Image-0.1-Edit-Turbo)
 Edit-Turbo is the distilled image-editing checkpoint. Its `model_index.json`
 declares `BooguImagePipeline`, so it uses the same native TI2I path as Edit.
 The upstream 1K hotfix is recommended for stable results and is pinned below
 to avoid silently loading the older checkpoint from the repository's default
 revision.
 ### Command
 ```bash
 vllm serve Boogu/Boogu-Image-0.1-Edit-Turbo \
  --omni \
  --revision hotfix-1k-20260708 \
  --port 8091
 ```
 ### Verification
 Edit an image with the distilled four-step settings:
 ```bash
 curl -s http://localhost:8091/v1/images/edits \
  -F model="Boogu/Boogu-Image-0.1-Edit-Turbo" \
  -F image="@input.png" \
  -F prompt="Change the style to a colored pencil drawing." \
  -F num_inference_steps=4 \
  -F guidance_scale=1.0 \
  -F seed=42 \
  | jq -r '.data[0].b64_json' | base64 -d > edited-turbo.png
 ```
 Or via chat completions:
 ```bash
 curl -s http://localhost:8091/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
        {"type": "text", "text": "Change the style to a colored pencil drawing."}
      ]}
    ],
    "extra_body": {
      "num_inference_steps": 4,
      "guidance_scale": 1.0,
      "seed": 42
    }
  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2- | base64 -d > edited-turbo.png
 ```
 ### Notes
 - **Pinned revision:** use `hotfix-1k-20260708`; upstream recommends the 1K
  hotfix over the 1.5K variant for more stable results.
 - **Recommended settings:** `num_inference_steps=4` and
  `guidance_scale=1.0`. Edit-Turbo is guidance-distilled, meaning the guidance
  behavior is already baked into the distilled checkpoint. A scale of `1.0`
  avoids applying additional classifier-free guidance; unlike the regular Edit
  checkpoint, it should not be increased to `5.0`.
 - **Reference images:** the same single-reference and `align_res` behavior as
  the regular Edit checkpoint applies.
 - **Known limitations:** the same single-GPU limitations as Base and Edit apply;
  CPU offload, Cache-DiT, and multi-GPU parallelism are not yet validated.
@@ -10,14 +10,22 @@ checkpoint for T2V and first-frame I2V. Every output includes synchronized
 48 kHz stereo audio.
 The raw [`Lightricks/LTX-2.5`](https://huggingface.co/Lightricks/LTX-2.5)
-repository is not directly loadable with `--model`. It supplies the official
+repository is not directly loadable with `--model`. It supplies the canonical
-upsampler and LoRA sidecars used by the Full/SFT two-stage pipeline, so accept
+Diffusion VAE checkpoint and the official upsampler and LoRA sidecars used by
-both model licenses and authenticate before first use.
+the Full/SFT two-stage pipeline, so accept both model licenses and authenticate
 before first use.
 The vLLM-Omni integration code is provided under the project's
 [Apache-2.0 license](../../LICENSE). The LTX-2.5 model weights are separate
 artifacts covered by the
 [LTX-2.x Community License](https://github.com/Lightricks/LTX-2/blob/main/LICENSE.md),
 not the vLLM-Omni license. Users are responsible for reviewing and complying
 with the model license before downloading or using the weights.
 ## Choose a pipeline
 | Pipeline | Mode | Default size | Schedule |
-|---|---|---:|---:|
+| --- | --- | ---: | ---: |
 | `LTX2Pipeline` | Full/SFT one-stage | 960x544 | 30 steps |
 | `LTX2TwoStagePipeline` | Full/SFT two-stage | 1920x1088 | 30 + 3 steps |
 | `LTX2DistilledOneStagePipeline` | Distilled one-stage | 960x544 | 8 steps |
@@ -28,6 +36,40 @@ latent upsampler, and run a three-step refinement stage. Select the class with
 `--model-class-name`; no `--task-type` flag is required. Supplying one initial
 image selects I2V, while omitting it selects T2V.
 ## Video decoder
 LTX-2.5 uses the canonical Native Diffusion VAE decoder by default. To opt in
 to the legacy convolutional VAE, set the startup-only stage-0 model extra
 `ltx2_use_conv_vae: true`. The choice applies to all requests and all four
 pipeline classes above. DiffVAE loads its NATTEN kernel from Hugging Face Hub
 during startup. `decode_timestep` and `decode_noise_scale` condition only the
 legacy ConvVAE and have no effect when DiffVAE is selected.
 For offline Python usage:
 ```python
 omni = Omni(
    model="Lightricks/LTX-2.5-Diffusers",
    model_class_name="LTX2Pipeline",
    stage_overrides='{"0":{"extras":{"ltx2_use_conv_vae":true}}}',
 )
 ```
 For online serving:
 ```bash
 vllm serve Lightricks/LTX-2.5-Diffusers \
  --omni \
  --model-class-name LTX2Pipeline \
  --stage-overrides '{"0":{"extras":{"ltx2_use_conv_vae":true}}}'
 ```
 Both decoders are untiled by default. Set `vae_use_tiling` for memory-saving
 serial tiling; DiffVAE tiles only above 80 frames or 768 pixels in either spatial
 dimension. DiffVAE also supports distributed VAE decode; see the
 [VAE Parallelism Guide](../../docs/user_guide/diffusion/parallelism/vae_parallelism.md).
 DiffVAE is decoder-only, so I2V still uses the convolutional VAE encoder.
 ## Prerequisites
 ```bash
@@ -42,7 +84,7 @@ Install matching vLLM and vLLM-Omni versions, and ensure `ffmpeg` and
 ## Hardware
 | GPU | Status | Recommended scope |
-|---|---|---|
+| --- | --- | --- |
 | NVIDIA B300 | Verified | All four canonical pipelines |
 | NVIDIA B200 or H200 | Capacity-based recommendation; not yet verified | All four canonical pipelines |
 | NVIDIA GB200 or GB300 | Capacity-based recommendation; not yet verified | All four canonical pipelines |
@@ -0,0 +1,82 @@
 # MiniMax H3 disaggregated text encoder
 This opt-in topology runs the Qwen3-VL text encoder as a vLLM stage and sends
 its hidden states and token-role metadata to an encoder-free diffusion stage.
 The standard MiniMax H3 recipes remain single-stage and continue to load the
 text encoder inside the diffusion pipeline.
 ## Start the server
 Choose the topology explicitly and load its deployment defaults:
 ```bash
 vllm-omni serve MiniMaxAI/MiniMax-H3 \
  --omni \
  --deploy-config vllm_omni/deploy/minimax_h3_disaggregated.yaml
 ```
 The default deployment assigns stage 0 to GPUs 0-1 with tensor parallel size
 2 and `max_num_seqs: 1`. Stage 1 uses GPUs 2-5 with diffusion tensor parallel
 size 1, Ulysses degree 4, and VAE patch parallel size 4. Adjust the
 `devices`, `tensor_parallel_size`, and stage 1 `parallel_config` values in a
 deployment override for the available hardware. Diffusion quantization,
 layerwise offload, distributed layerwise offload, VAE parallelism, and USP
 settings use the same stage 1 options documented in [MiniMax-H3.md](MiniMax-H3.md).
 For example, this five-GPU topology assigns one GPU to the encoder and four
 to the diffusion stage. `--stage-overrides` keeps placement and parallelism
 scoped to the owning stage rather than broadcasting an override to both:
 ```bash
 CUDA_VISIBLE_DEVICES=0,1,2,3,4 \
 vllm-omni serve MiniMaxAI/MiniMax-H3 \
  --omni \
  --deploy-config vllm_omni/deploy/minimax_h3_disaggregated.yaml \
  --stage-overrides '{"0":{"devices":"0","tensor_parallel_size":1},"1":{"devices":"1,2,3,4","tensor_parallel_size":1,"ulysses_degree":4,"vae_patch_parallel_size":4}}'
 ```
 For memory-constrained deployments, start from the CPU-offload or distributed
 layerwise-offload profiles in [MiniMax-H3.md](MiniMax-H3.md). Apply memory and
 quantization options only to Stage 1 with `--stage-overrides`; retain the
 encoder's BF16 configuration and the video/audio VAEs' FP32 precision. Select
 one offload strategy per deployment:
 ```bash
 # Stage 1 model-level CPU offload.
 --stage-overrides '{"1":{"enable_cpu_offload":true}}'
 # Stage 1 distributed layerwise offload. Tune resident layers for available RAM.
 --stage-overrides '{"1":{"enable_distributed_layerwise_offload":true,"dlo_use_allgather":false,"dlo_resident_layers":20}}'
 # Stage 1 online FP8 quantization of the DiT only.
 --stage-overrides '{"1":{"diffusion_quantization_config":"{\"transformer\":{\"method\":\"fp8\"}}"}}'
 ```
 The Stage 1 VAE patch-parallel options remain independent of offload and
 quantization. See [MiniMax-H3.md](MiniMax-H3.md) for memory requirements and
 hardware-qualified profiles before combining these options.
 Stage 1 sets `model_loaded.text_encoder: false`; it must not load or download
 text-encoder weights. This H3 topology explicitly keeps its single-replica
 diffusion stage inline, avoiding serialization of decoded video through a
 subprocess. It expects both stages to run in one deployment; cross-node payload
 transport is outside this configuration.
 The `/v1/videos` request schema and `extra_params.task` values (`t2va`,
 `fl2va`, and `ref2va`) are unchanged from the single-stage recipe.
 ## Turbo LoRA
 MiniMax-H3 Turbo uses five sigma points, `flow_shift=6`, and
 `audio_flow_shift=3`. Start Turbo deployments with the dedicated defaults so
 requests that omit sampling controls do not inherit the 50-step base schedule:
 ```bash
 vllm-omni serve MiniMaxAI/MiniMax-H3 \
  --omni \
  --lora-path /path/to/MiniMax-H3-Turbo \
  --deploy-config vllm_omni/deploy/minimax_h3_disaggregated_turbo.yaml
 ```
 The base deployment intentionally retains 50 inference steps for non-LoRA
 quality. Turbo LoRA supports T2VA and FL2VA requests, not Ref2VA.
@@ -69,11 +69,12 @@ reference-video preparation and MP4 output.
 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,
+path automatically at runtime. The server-owned parallel converter accepts
-and RGB per-channel contiguous layout use direct planar frames; other inputs
+supported frame shapes and dtypes with either per-channel-contiguous or strided
-fall back to the legacy path before the PyAV container is opened. No CLI flag,
+RGB planes, including interleaved arrays materialized by output transport.
-model declaration, or user configuration is required. Streaming fMP4 output is
+Standalone callers without a parallel converter retain the legacy fallback for
-unchanged.
+strided planes. 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
@@ -860,6 +861,129 @@ fixed HBM usage. Model-level and standard layerwise offload remain unsupported.
 The five requested sigma points produce the four denoiser evaluations expected
 by the Turbo artifact.
 ### FlashGen native LoRA
 The FlashGen 4-step T2VA artifact uses the native MiniMax-H3 module layout and
 declares its distilled sigma schedule in safetensors metadata. It is published on
 [ModelScope](https://modelscope.cn/models/FlashGen/Minimax-H3-4step-lora-flashgen):
 ```text
 FlashGen/Minimax-H3-4step-lora-flashgen/minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors
 ```
 Download only that file:
 ```bash
 python -m pip install modelscope
 export FLASHGEN_DIR=/path/to/minimax-h3-flashgen-lora
 export FLASHGEN_FILE=minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors
 modelscope download FlashGen/Minimax-H3-4step-lora-flashgen \
  --local_dir "${FLASHGEN_DIR}" \
  --include "${FLASHGEN_FILE}"
 export FLASHGEN_LORA="${FLASHGEN_DIR}/${FLASHGEN_FILE}"
 ```
 Start from a non-offloaded or DLO FL2VA server command and add
 `--task-type fl2va --lora-backend peft --lora-path "${FLASHGEN_LORA}"`.
 Each request must use T2VA and the distilled interval-count contract:
 ```bash
 -F 'num_inference_steps=4' \
 -F 'extra_params={"task":"t2va","duration":5.2}' \
 -F "lora={\"name\":\"h3-flashgen-v1.0\",\"path\":\"${FLASHGEN_LORA}\",\"scale\":1.0}"
 ```
 This path rejects Ref2VA and checkpoints that already pin `base_schedule` in
 `model_index.json`. The adapter metadata carries
 `base_schedule=1.0,0.7,0.4,0.15,0.0`, so `num_inference_steps=4` means four
 denoiser evaluations, not five sigma points. Request-mode generation may omit
 the field and take the count from the adapter schedule; `--step-execution`
 requires it explicitly, because the step scheduler reads the total step count
 off the request at admission, before the adapter schedule is known.
 DLO is supported in request-mode generation on the same terms as the Turbo
 adapter: the request-switchable LoRA A/B buffers stay resident on the
 accelerator while DLO streams only the base blocks, so budget for that
 additional fixed HBM usage. The native artifact is rank 64 over 259 target
 modules, and its packed `qkv_proj` and `fc1` layers reuse the full-input A
 tensor per slice while B carries slice-local output rows, so the resident
 footprint exceeds the on-disk payload; measure it for your parallel layout
 rather than assuming the checkpoint size. Pure Ulysses replicates the adapter
 on every rank, while DiT tensor parallelism shards the B buffers. Model-level
 and standard layerwise offload remain unsupported, and `--step-execution`
 cannot be combined with `--enable-distributed-layerwise-offload`.
 To validate a deployment, post the same fixed-seed T2VA request twice with the
 adapter and twice without it, then compare the four output digests. The adapter
 is bound and deterministic when each pair matches internally and the two pairs
 differ from each other.
 ### FastH3 adapter
 [FastH3](https://haoailab.com/blogs/fasth3-preview/) is FastVideo's four-step
 DMD2 student of H3-Base. It reuses H3's text encoder, VAEs, tokenizers, and
 schedulers unchanged, replacing 49 denoiser evaluations with four. It is fused
 into the checkpoint at load time rather than switched per request, because it
 carries full-rank deltas that no LoRA layer can express.
 The bundle publishes four variants, so download one and point `--lora-path` at
 it; the repository root is refused rather than guessed at:
 ```bash
 export FASTH3_DIR=/path/to/fasth3
 hf download FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA \
  dense-datafree/adapter_model.safetensors --local-dir "${FASTH3_DIR}"
 export FASTH3_LORA="${FASTH3_DIR}/dense-datafree/adapter_model.safetensors"
 ```
 Add `--task-type fl2va --lora-path "${FASTH3_LORA}"` to a non-offloaded server
 command. T2VA is served by the FL2VA partition, so `--task-type fl2va` is
 correct even though FastH3 preview v1 distills T2VA only. Because the adapter is
 fused, `--lora-backend` does not apply and a request carrying a `lora=` field is
 rejected rather than served without the adapter it asked for.
 ```bash
 -F 'num_inference_steps=4' \
 -F 'extra_params={"task":"t2va","duration":4.4}'
 ```
 Requests must ask for `num_inference_steps=4` and `task=t2va`: the release's five
 sigma points bound four denoiser evaluations, and that count is what the step
 scheduler admits a request on. The server denoises on the release's own ladder,
 keeping H3's per-modality shifts at the checkpoint values, so a request that
 overrides `flow_shift` or `audio_flow_shift` is rejected - it would sample the
 student at noise levels it was never distilled at.
 Only a release that identifies itself as FastH3 is fused; any other
 `fastvideo-lora-v2` adapter stays on the dynamic LoRA route. A claimed artifact
 is then held to its own metadata: one that misdeclares its tensor counts or
 leaves a transformer block unedited is refused at startup instead of serving
 mostly base H3 weights on a four-step schedule. Offload is refused for the same
 reason - `--enable-cpu-offload`, `--enable-layerwise-offload` and
 `--enable-distributed-layerwise-offload` all bypass the fusion, so they fail fast.
 > [!NOTE]
 > Only the **Dense / Data-Free** variant is supported today. The three VSA
 > variants carry `.set_weight` compression gates for modules that do not exist
 > in vLLM-Omni's dense H3 attention, and the release states dense attention is
 > not a supported substitute for them, so they are refused at startup rather
 > than silently run wrong. Video Sparse Attention for H3's packed
 > `[text | cond | audio | video]` sequence is follow-up work.
 Measured on 8x NVIDIA B300 with USP8, VAE patch-parallel 8, `TRTLLM_ATTN`, at
 1344x768, 4.4 s, seed 1101, one warmup excluded and two runs recorded:
 | Adapter | Steps | End-to-end | Diffusion engine |
 | --- | ---: | ---: | ---: |
 | none (base H3) | 50 | 25.8 / 26.4 s | 16.22 / 16.28 s |
 | FastH3 Dense | 4 (5 sigma points) | 11.7 / 11.8 s | 2.37 / 2.36 s |
 The denoising speedup is 6.9x. End-to-end is 2.2x because text encoding, VAE
 decoding and muxing are a fixed cost that dominates a clip this short; longer
 generations move the end-to-end figure toward the denoising one. Fusing the
 adapter does not measurably change startup: weight loading took 77.3 s with it
 against 85.8 s without.
 ## Key parameters
 | Parameter | Recommended value | Notes |
@@ -18,11 +18,11 @@ examples in this repository.
 Qwen3-TTS supports three task types, each backed by a dedicated model checkpoint:
-| Task Type     | Model                                    | Description                                                  |
+| Task Type     | Model                                    | Description                                                   |
-| ------------- | ---------------------------------------- | ------------------------------------------------------------ |
+| ------------- | ---------------------------------------- | ------------------------------------------------------------- |
-| `CustomVoice` | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice`  | Predefined speaker voices with optional style/emotion control |
+| `CustomVoice` | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice`   | Predefined speaker voices with optional style/emotion control |
-| `VoiceDesign` | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`  | Generate speech from a natural language voice description     |
+| `VoiceDesign` | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`   | Generate speech from a natural language voice description     |
-| `Base`        | `Qwen/Qwen3-TTS-12Hz-1.7B-Base`         | Voice cloning from reference audio + transcript              |
+| `Base`        | `Qwen/Qwen3-TTS-12Hz-1.7B-Base`          | Voice cloning from reference audio + transcript               |
 Smaller 0.6B variants are also available for `CustomVoice` and `Base`.
@@ -166,6 +166,15 @@ python examples/offline_inference/text_to_speech/qwen3_tts/end2end.py --query-ty
 - Key flags: `--omni` is required. `--deploy-config` points to the bundled two-stage pipeline config.
 - Async chunking: Enabled by default in `qwen3_tts.yaml` for streaming-friendly first-audio latency. Raw audio streaming requires `stream=true`, `stream_format="audio"`, and `response_format="pcm"`.
 - Task/model matching: Each task type requires its matching model checkpoint. Using a CustomVoice model for a Base (voice clone) request will fail.
 - Base codec termination: Base requests without an explicit `max_new_tokens`
  use a text-scaled safety ceiling (at least 192 codec frames and no more than
  the configured model limit). If the Talker reaches that ceiling without
  codec EOS, non-streaming serving discards the incomplete audio and retries
  once with a fresh seed; an explicit `seed` or `max_new_tokens` disables the
  retry. SSE and WebSocket clients receive structured errors and must discard
  previously emitted audio when the error contains `"action":"discard"`;
  raw PCM streams terminate with a connection error after any partial bytes
  already sent.
 - Known limitations: The server serves one model variant at a time. To switch task types (e.g., CustomVoice to Base), restart the server with the corresponding model.
 ## Hardware Support
@@ -13,6 +13,8 @@ transformers >= 5.10.1, < 5.15
 av>=14.0.0
 omegaconf>=2.3.0
 diffusers==0.40.0
 # Hub kernels, including LTX-2.5 NATTEN
 kernels==0.14.1
 safetensors>=0.8.0
 accelerate==1.12.0
 soundfile>=0.13.1
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Benchmark legacy and automatic MP4 response encoding on synthetic input."""
 from __future__ import annotations
@@ -11,13 +11,16 @@ import statistics
 import time
 from collections.abc import Callable
 from dataclasses import dataclass
 from functools import partial
 from pathlib import Path
 from typing import cast
 import numpy as np
 from vllm_omni.entrypoints.openai.video_api_utils import (
    _encode_video_bytes,
    _encode_video_bytes_legacy,
    _PlanarFrameConverter,
 )
@@ -42,12 +45,19 @@ def _build_inputs(
    fps: int,
    audio_sample_rate: int,
    seed: int,
    input_layout: str = "planar-backed",
 ) -> 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)):
+    if input_layout == "interleaved":
        video = np.ascontiguousarray(video)
        if not video.flags.c_contiguous:
            raise RuntimeError("synthetic video is not C-contiguous")
    elif input_layout == "planar-backed" and not all(video[..., channel].flags.c_contiguous for channel in range(3)):
        raise RuntimeError("synthetic video does not expose contiguous channel planes")
    elif input_layout != "planar-backed":
        raise ValueError(f"unsupported input layout: {input_layout}")
    sample_count = max(1, round(frames / fps * audio_sample_rate))
    timeline = np.arange(sample_count, dtype=np.float32) / audio_sample_rate
@@ -86,8 +96,8 @@ def _measure(
 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]
+    wall_values = [cast(float, record["wall_ms"]) for record in selected]
-    cpu_values = [float(record["process_cpu_ms"]) for record in selected]
+    cpu_values = [cast(float, record["process_cpu_ms"]) for record in selected]
    return {
        "runs": len(selected),
        "wall_ms": {
@@ -113,6 +123,16 @@ def _parse_args() -> argparse.Namespace:
    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(
        "--input-layout",
        choices=("planar-backed", "interleaved"),
        default="planar-backed",
    )
    parser.add_argument(
        "--frame-conversion-workers",
        type=_positive_int,
        default=8,
    )
    parser.add_argument("--output-json", type=Path)
    args = parser.parse_args()
    if args.warmup < 0:
@@ -129,33 +149,41 @@ def main() -> None:
        fps=args.fps,
        audio_sample_rate=args.audio_sample_rate,
        seed=args.seed,
        input_layout=args.input_layout,
    )
    baseline = BenchmarkVariant("legacy", _encode_video_bytes_legacy)
-    candidate = BenchmarkVariant("automatic", _encode_video_bytes)
+    converter = _PlanarFrameConverter(max_workers=args.frame_conversion_workers)
    candidate = BenchmarkVariant(
        "automatic",
        partial(_encode_video_bytes, frame_converter=converter),
    )
-    for _ in range(args.warmup):
+    try:
-        for variant in (baseline, candidate):
+        for _ in range(args.warmup):
-            _measure(
+            for variant in (baseline, candidate):
-                variant,
+                _measure(
-                video=video,
+                    variant,
-                audio=audio,
+                    video=video,
-                fps=args.fps,
+                    audio=audio,
-                audio_sample_rate=args.audio_sample_rate,
+                    fps=args.fps,
-            )
+                    audio_sample_rate=args.audio_sample_rate,
                )
-    records: list[dict[str, object]] = []
+        records: list[dict[str, object]] = []
-    for round_index in range(args.rounds):
+        for round_index in range(args.rounds):
-        order = (baseline, candidate) if round_index % 2 == 0 else (candidate, baseline)
+            order = (baseline, candidate) if round_index % 2 == 0 else (candidate, baseline)
-        for variant in order:
+            for variant in order:
-            _, record = _measure(
+                _, record = _measure(
-                variant,
+                    variant,
-                video=video,
+                    video=video,
-                audio=audio,
+                    audio=audio,
-                fps=args.fps,
+                    fps=args.fps,
-                audio_sample_rate=args.audio_sample_rate,
+                    audio_sample_rate=args.audio_sample_rate,
-            )
+                )
-            record["round"] = round_index + 1
+                record["round"] = round_index + 1
-            records.append(record)
+                records.append(record)
    finally:
        converter.shutdown()
    output_hashes = {str(record["output_sha256"]) for record in records}
    if len(output_hashes) != 1:
@@ -165,8 +193,8 @@ def main() -> None:
    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))
+        baseline_ms = cast(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))
+        candidate_ms = cast(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)
@@ -180,6 +208,8 @@ def main() -> None:
            "warmup": args.warmup,
            "rounds": args.rounds,
            "seed": args.seed,
            "input_layout": args.input_layout,
            "frame_conversion_workers": args.frame_conversion_workers,
            "video_shape": list(video.shape),
            "video_strides": list(video.strides),
            "codec_options": {"preset": "ultrafast", "threads": "0"},
@@ -279,17 +279,79 @@ def test_unmeasured_tpot_stays_missing_after_tokenizer_fallback():
    assert metrics.request_goodput == 0.0
-def test_measured_zero_itl_is_not_treated_as_missing():
+def test_zero_itl_does_not_create_zero_tpot():
    output = _make_output(100, output_tokens=3)
    output.itl = [0.0, 0.0]
    output.duplex_session_metrics = {"mean_ttft_ms": 100.0}
    metrics = _calculate_test_metrics([output], {"tpot": 1.0})
-    assert (metrics.num_tpot_samples, metrics.num_itl_samples) == (1, 2)
+    assert (metrics.num_tpot_samples, metrics.num_itl_samples) == (0, 2)
-    assert metrics.mean_tpot_ms == 0.0
+    assert math.isnan(metrics.mean_tpot_ms)
    assert metrics.mean_itl_ms == 0.0
-    assert metrics.request_goodput == 1.0
+    assert metrics.request_goodput == 0.0
 def test_stage_output_tokens_without_client_timing_omit_tpot(capsys):
    output = _make_output(100, output_tokens=5)
    output.itl = []
    output.text_latency = 0.0
    output.ttft = 0.1
    metrics, _ = calculate_metrics(
        input_requests=[],
        outputs=[output],
        dur_s=1.0,
        tokenizer=None,
        selected_percentiles=[50.0, 99.0],
        goodput_config_dict={},
        task_type=TaskType.GENERATION,
        selected_percentile_metrics=["tpot"],
        max_concurrency=None,
        request_rate=float("inf"),
        benchmark_duration=1.0,
    )
    printed = capsys.readouterr().out
    assert metrics.num_tpot_samples == 0
    assert math.isnan(metrics.mean_tpot_ms)
    assert "Time per Output Token" not in printed
    assert "Mean TPOT" not in printed
 def test_consistent_client_latency_can_supply_tpot_fallback():
    output = _make_output(100, output_tokens=10)
    output.itl = []
    output.ttft = 0.1
    output.text_latency = 1.0
    metrics = _calculate_test_metrics([output])
    assert metrics.num_tpot_samples == 1
    assert metrics.mean_tpot_ms == pytest.approx(100.0)
 def test_single_token_responses_do_not_report_zero_tpot(capsys):
    output = _make_output(100, output_tokens=1)
    metrics, _ = calculate_metrics(
        input_requests=[],
        outputs=[output],
        dur_s=1.0,
        tokenizer=None,
        selected_percentiles=[50.0],
        goodput_config_dict={},
        task_type=TaskType.GENERATION,
        selected_percentile_metrics=["tpot"],
        max_concurrency=None,
        request_rate=float("inf"),
        benchmark_duration=1.0,
    )
    printed = capsys.readouterr().out
    assert metrics.num_tpot_samples == 0
    assert math.isnan(metrics.mean_tpot_ms)
    assert "Time per Output Token" not in printed
 def test_duplex_goodput_does_not_pair_measurements_from_different_requests():
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 Unit tests for patch.py
@@ -9,22 +9,98 @@ import asyncio
 import base64
 import json
 import time
 from argparse import Namespace
 from types import SimpleNamespace
 import pytest
 from pytest_mock import MockerFixture
 from vllm.benchmarks.lib.endpoint_request_func import RequestFuncInput
 from vllm_omni.benchmarks.data_modules.seed_tts_dataset import (
    SeedTTSSampleRequest,
    SeedTTSTextSampleRequest,
 )
 from vllm_omni.benchmarks.patch.patch import (
    MixRequestFuncOutput,
    _apply_stage0_token_timings,
    _attach_seed_tts_to_request_func_input,
    async_request_openai_chat_omni_completions,
    async_request_openai_realtime_duplex,
    should_request_stage_metrics,
 )
 from vllm_omni.experimental.fullduplex.client import RealtimeEventCollector
 pytestmark = [pytest.mark.core_model, pytest.mark.benchmark, pytest.mark.cpu]
 def _seed_tts_request_func_input() -> RequestFuncInput:
    return RequestFuncInput(
        model="test-model",
        model_name="test-model",
        prompt="target text",
        api_url="http://test.com/v1/chat/completions",
        prompt_len=2,
        output_len=20,
        extra_body={"modalities": ["text", "audio"]},
    )
 def test_seed_tts_chat_request_carries_reference_audio_once() -> None:
    reference_audio_url = "data:audio/wav;base64,AAAA"
    sample = SeedTTSSampleRequest(
        prompt="target text",
        prompt_len=2,
        expected_output_len=20,
        multi_modal_data=None,
        seed_tts_system_prompt="Clone the supplied voice.",
        seed_tts_speech_extra={
            "ref_audio": reference_audio_url,
            "ref_text": "reference text",
            "task_type": "Base",
        },
    )
    request_func_input = _seed_tts_request_func_input()
    _attach_seed_tts_to_request_func_input(sample, request_func_input)
    assert request_func_input.omni_chat_messages == [
        {
            "role": "system",
            "content": [{"type": "text", "text": "Clone the supplied voice."}],
        },
        {"role": "user", "content": [{"type": "text", "text": "target text"}]},
    ]
    assert request_func_input.extra_body["ref_audio"] == reference_audio_url
    serialized_request = json.dumps(
        {"messages": request_func_input.omni_chat_messages, **request_func_input.extra_body},
    )
    assert serialized_request.count(reference_audio_url) == 1
 def test_seed_tts_text_chat_messages_do_not_add_reference_audio() -> None:
    sample = SeedTTSTextSampleRequest(
        prompt="target text",
        prompt_len=2,
        expected_output_len=20,
        multi_modal_data=None,
        seed_tts_system_prompt="Use the configured voice.",
        seed_tts_speech_extra=None,
    )
    request_func_input = _seed_tts_request_func_input()
    _attach_seed_tts_to_request_func_input(sample, request_func_input)
    assert getattr(request_func_input, "seed_tts_row", False) is True
    assert request_func_input.omni_chat_messages == [
        {
            "role": "system",
            "content": [{"type": "text", "text": "Use the configured voice."}],
        },
        {"role": "user", "content": [{"type": "text", "text": "target text"}]},
    ]
    assert "ref_audio" not in request_func_input.extra_body
 class MockResponse:
    """Mock aiohttp response for testing"""
@@ -102,7 +178,24 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
                received_at_s=now + 0.03,
            )
            self.events.add(
-                {"type": "response.done", "response": {"id": response_id}},
+                {
                    "type": "response.done",
                    "response": {
                        "id": response_id,
                        "metadata": {
                            "vllm_omni": {
                                "stage_metrics": {
                                    "0": {
                                        "num_tokens_out": 3,
                                        "vllm_ttft_ms": 20.0,
                                        "vllm_tpot_ms": 10.0,
                                        "vllm_itls_ms": [10.0, 10.0],
                                    }
                                }
                            }
                        },
                    },
                },
                received_at_s=now + 0.04,
            )
@@ -143,6 +236,7 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
    assert client.configure_kwargs["native_duplex"] is False
    assert client.configure_kwargs["extra_body"] == {
        "ref_audio": "data:audio/wav;base64,AAAA",
        "return_stage_metrics": True,
    }
    assert client.sent == [
        event
@@ -167,6 +261,10 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
    assert output.audio_ttfp > output.ttft
    assert output.audio_rtf > 0
    assert output.latency > 0
    assert output.output_tokens == 12
    assert output.itl == [0.01, 0.01] * 4
    assert output.text_latency == pytest.approx(output.ttft + 0.08)
    assert output.tpot_measured is True
    assert output.tts_turn_pcm_bytes == [b"\x00\x00" * 2400] * 4
    assert output.tts_output_pcm_bytes == b"\x00\x00" * 9600
    session_id = output.duplex_request_metrics[0]["session_id"]
@@ -198,11 +296,160 @@ async def test_seed_tts_realtime_duplex_exports_per_request_metrics(monkeypatch)
    }
 def test_stage0_token_timings_use_weighted_tpot_when_itls_are_incomplete():
    output = MixRequestFuncOutput(ttft=0.1)
    measured = _apply_stage0_token_timings(
        output,
        [
            {"output_token_count": 3, "itls_ms": [], "tpot_ms": 10.0},
            {"output_token_count": 2, "itls_ms": [], "tpot_ms": 40.0},
        ],
        expected_output_tokens=5,
    )
    assert measured is True
    assert output.itl == []
    assert output.tpot_measured is True
    assert (output.text_latency - output.ttft) / 4 == pytest.approx(0.02)
 def test_stage0_zero_itls_fall_back_to_positive_tpot():
    output = MixRequestFuncOutput(ttft=0.1)
    measured = _apply_stage0_token_timings(
        output,
        [{"output_token_count": 3, "itls_ms": [0.0, 0.0], "tpot_ms": 25.0}],
        expected_output_tokens=3,
    )
    assert measured is True
    assert output.itl == []
    assert output.tpot_measured is True
    assert (output.text_latency - output.ttft) / 2 == pytest.approx(0.025)
 def test_stage0_token_count_without_timing_is_not_measured():
    output = MixRequestFuncOutput(ttft=0.1)
    measured = _apply_stage0_token_timings(
        output,
        [{"output_token_count": 3, "itls_ms": [], "tpot_ms": None}],
        expected_output_tokens=3,
    )
    assert measured is False
    assert output.itl == []
    assert output.text_latency == output.ttft
    assert output.tpot_measured is False
 def create_sse_chunk(data_dict):
    """Helper to create SSE formatted chunk"""
    return f"data: {json.dumps(data_dict)}\n\n".encode()
 def test_chat_text_timing_metrics_request_stage_metrics():
    args = Namespace(
        backend="openai-chat-omni",
        percentile_metrics="ttft,tpot,itl,e2el",
        print_stage=False,
        extra_body={},
    )
    assert should_request_stage_metrics(args) is True
@pytest.mark.asyncio
 async def test_bundled_first_text_chunk_uses_stage0_token_timings(mocker: MockerFixture):
    """Engine timings recover TPOT when every text token arrives together."""
    request_input = RequestFuncInput(
        model="test-model",
        model_name="test-model",
        prompt="test prompt",
        api_url="http://test.com/v1/chat/completions",
        prompt_len=10,
        output_len=20,
    )
    chunks = [
        create_sse_chunk(
            {
                "choices": [{"delta": {"content": "ABCD"}}],
                "modality": "text",
                "usage": {"prompt_tokens": 10, "completion_tokens": 4, "total_tokens": 14},
                "metrics": {
                    "stage_metrics": {
                        "0": {
                            "num_tokens_out": 4,
                            "vllm_itls_ms": [10.0, 11.0, 12.0],
                            "vllm_tpot_ms": 11.0,
                        }
                    }
                },
            }
        ),
        b"data: [DONE]\n\n",
    ]
    mock_response = MockResponse(200, chunks)
    mock_session = mocker.AsyncMock()
    mock_session.post = mocker.MagicMock(return_value=mock_response)
    output = await async_request_openai_chat_omni_completions(request_input, mock_session)
    assert output.success is True
    assert output.output_tokens == 4
    assert output.itl == pytest.approx([0.010, 0.011, 0.012])
    assert output.text_latency - output.ttft == pytest.approx(0.033)
    assert output.tpot_measured is True
@pytest.mark.asyncio
 async def test_positive_client_text_timings_take_precedence_over_stage0(mocker: MockerFixture):
    request_input = RequestFuncInput(
        model="test-model",
        model_name="test-model",
        prompt="test prompt",
        api_url="http://test.com/v1/chat/completions",
        prompt_len=10,
        output_len=20,
    )
    chunks = [
        create_sse_chunk(
            {
                "choices": [{"delta": {"content": "A"}}],
                "modality": "text",
                "usage": {"prompt_tokens": 10, "completion_tokens": 1, "total_tokens": 11},
            }
        ),
        create_sse_chunk(
            {
                "choices": [{"delta": {"content": "B"}}],
                "modality": "text",
                "usage": {"prompt_tokens": 10, "completion_tokens": 2, "total_tokens": 12},
                "metrics": {
                    "stage_metrics": {
                        "0": {
                            "num_tokens_out": 2,
                            "vllm_itls_ms": [1000.0],
                            "vllm_tpot_ms": 1000.0,
                        }
                    }
                },
            }
        ),
        b"data: [DONE]\n\n",
    ]
    mock_response = MockResponse(200, chunks, delay_between_chunks=0.01)
    mock_session = mocker.AsyncMock()
    mock_session.post = mocker.MagicMock(return_value=mock_response)
    output = await async_request_openai_chat_omni_completions(request_input, mock_session)
    assert output.success is True
    assert len(output.itl) == 1
    assert 0.0 < output.itl[0] < 0.1
 # ============================================================================
 # output_tokens Tests
 # ============================================================================
@@ -1,3 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for Daily-Omni ``--daily-omni-pack-mode`` packing recipes.
 The ``minicpm-interleave`` mode is an accuracy-critical protocol: MiniCPM-o only reaches
@@ -27,6 +30,8 @@ _MODULE_NAME = "vllm_omni.benchmarks.data_modules.daily_omni_dataset"
 if _MODULE_NAME not in sys.modules:
    _spec = importlib.util.spec_from_file_location(_MODULE_NAME, _MODULE_PATH)
    if _spec is None or _spec.loader is None:
        raise ImportError(f"Cannot load {_MODULE_NAME} from {_MODULE_PATH}")
    _mod = importlib.util.module_from_spec(_spec)
    sys.modules[_MODULE_NAME] = _mod
    _spec.loader.exec_module(_mod)
@@ -87,7 +92,9 @@ def _stub_extract(monkeypatch, dataset: DailyOmniDataset, counter: list[int]) ->
    def fake_extract(video_path, *, audio_path, include_audio, **_kw):
        counter.append(1)
        frames = [Image.new("RGB", (16, 16), color=(i, i, i)) for i in range(_NUM_FRAMES)]
-        segments = [np.zeros(1600, dtype=np.float32) for _ in range(_NUM_FRAMES)] if include_audio else []
+        segments: list[np.ndarray] = (
            [np.zeros(1600, dtype=np.float32) for _ in range(_NUM_FRAMES)] if include_audio else []
        )
        return frames, segments
    monkeypatch.setattr(dataset, "_extract_minicpm_frame_audio_segments", fake_extract)
@@ -585,7 +585,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
@pytest.mark.asyncio
@pytest.mark.parametrize(
-    ("request_metrics", "expected_itl", "expected_text_latency"),
+    ("request_metrics", "expected_itl", "expected_text_latency", "expected_tpot_measured"),
    [
        (
            [
@@ -594,6 +594,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
            ],
            [0.01, 0.02, 0.03],
            0.16,
            True,
        ),
        (
            [
@@ -602,6 +603,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
            ],
            [],
            0.18,
            True,
        ),
        (
            [
@@ -610,6 +612,7 @@ async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Pa
            ],
            [],
            0.1,
            False,
        ),
    ],
 )
@@ -619,6 +622,7 @@ async def test_adapter_reports_exact_or_weighted_token_timing(
    request_metrics: list[dict[str, object]],
    expected_itl: list[float],
    expected_text_latency: float,
    expected_tpot_measured: bool,
 ):
    case, options = _case(tmp_path), _session_options(tmp_path)
    request = _request(case, options)
@@ -643,7 +647,7 @@ async def test_adapter_reports_exact_or_weighted_token_timing(
    assert output.success
    assert output.itl == expected_itl
    assert output.text_latency == pytest.approx(expected_text_latency)
-    assert output.tpot_measured is True
+    assert output.tpot_measured is expected_tpot_measured
 def test_batch_finalization_publishes_only_measured_results(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
@@ -1,3 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for SeedTTSTextDataset, SeedTTSTextSampleRequest, SeedTTSDesignDataset,
 and SeedTTSDesignSampleRequest.
@@ -234,34 +237,6 @@ def test_seed_tts_design_dataset_rejects_missing_description(seed_tts_design_roo
        assert req.seed_tts_utterance_id == "ok"
 def test_attach_sets_seed_tts_row_even_without_extra_body():
    """seed_tts_row=True must be set for SeedTTSTextSampleRequest (no extra body)."""
    from vllm_omni.benchmarks.data_modules.seed_tts_dataset import SeedTTSTextSampleRequest
    req = SeedTTSTextSampleRequest(
        prompt="hello world",
        prompt_len=2,
        expected_output_len=100,
        multi_modal_data=None,
        request_id="test-0",
        seed_tts_speech_extra=None,
        seed_tts_ref_wav_path="",
    )
    assert req.seed_tts_speech_extra is None
    assert req.seed_tts_ref_wav_path == ""
    # The fix ensures that even with speech_extra=None, the function
    # sets seed_tts_row=True. We verify the source code has the fix.
    import inspect
    import vllm_omni.benchmarks.patch.patch as patch_mod
    src = inspect.getsource(patch_mod._attach_seed_tts_to_request_func_input)
    # seed_tts_row must be set BEFORE the 'if not ex: return' check
    row_pos = src.index("seed_tts_row")
    not_ex_pos = src.index("if not ex:")
    assert row_pos < not_ex_pos, "seed_tts_row must be set before 'if not ex: return'"
 def test_seed_tts_whisper_transcribe_passes_attention_mask(monkeypatch):
    from vllm_omni.benchmarks.data_modules import seed_tts_eval
@@ -34,12 +34,18 @@ from vllm_omni.config.stage_config import (
    build_stage_runtime_overrides,
    load_deploy_config,
    merge_pipeline_deploy,
    normalize_pipeline_cli_overrides,
    pipeline_cfg_resolver,
 )
 from vllm_omni.engine.arg_utils import SHARED_FIELDS, internal_blacklist_keys
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 SINGLE_STAGE_PIPE_CFG = (StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),)
 # Validation strings to match against in post init
 NO_STAGES_MATCH_STR = "no stages"
 NO_TERMINAL_STAGE_MATCH_STR = "No terminal stage"
@pytest.fixture(autouse=True)
 def _stable_test_platform(monkeypatch):
@@ -429,6 +435,20 @@ class TestStageConfigFactory:
        assert cfg["final_output"] is True
        assert cfg["final_output_type"] == "image"
    @pytest.mark.parametrize(
        "model_class_name",
        [
            "MiniMaxH3Pipeline",
            "WanPipeline",
            "LTX2Pipeline",
            "Cosmos3OmniDiffusersPipeline",
        ],
    )
    def test_default_video_diffusion_output_type(self, model_class_name):
        configs = StageConfigFactory.create_default_diffusion({"model_class_name": model_class_name})
        assert configs[0]["final_output_type"] == "video"
    def test_default_diffusion_with_parallel_config(self):
        """Test diffusion config calculates devices from parallel_config."""
@@ -603,6 +623,7 @@ class TestPipelineDiscovery:
        p = PipelineConfig(
            model_type="custom_collide",
            hf_architectures=("SomeCollidingArch",),
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        assert p.hf_architectures == ("SomeCollidingArch",)
@@ -624,33 +645,63 @@ class TestStagePipelineConfig:
 class TestPipelineConfigNew:
-    def test_frozen(self):
+    def test_simple_init(self):
-        p = PipelineConfig(model_type="t", model_arch="A")
+        # Ensure pipeline config validates in post init.
-        with pytest.raises(AttributeError):
+        PipelineConfig(
-            p.model_type = "changed"
+            model_type="t",
            model_arch="A",
            stages=SINGLE_STAGE_PIPE_CFG,
        )
-    def test_validate_valid(self):
+    def test_frozen(self):
        p = PipelineConfig(
            model_type="t",
            model_arch="A",
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        with pytest.raises(AttributeError):
            p.model_type = "changed"
    def test_validate_no_stages(self):
        with pytest.raises(ValueError, match=NO_STAGES_MATCH_STR):
            PipelineConfig(model_type="t", model_arch="A")
    def test_validate_no_terminal_stage(self):
        """A pipeline with no ``final_output`` stage can never emit a result."""
        with pytest.raises(ValueError, match=NO_TERMINAL_STAGE_MATCH_STR):
            PipelineConfig(
                model_type="t",
                model_arch="A",
                stages=(
                    StagePipelineConfig(stage_id=0, model_stage="a"),
                    StagePipelineConfig(stage_id=1, model_stage="b", input_sources=(0,)),
                ),
            )
    def test_pipeline_can_have_final_output_in_any_stage(self):
        """Ensure any stage may carry ``final_output``."""
        PipelineConfig(
            model_type="t",
            model_arch="A",
            stages=(
-                StagePipelineConfig(stage_id=0, model_stage="a"),
+                StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),
                StagePipelineConfig(stage_id=1, model_stage="b", input_sources=(0,)),
            ),
        )
        assert p.validate() == []
    def test_validate_no_stages(self):
        p = PipelineConfig(model_type="t", model_arch="A")
        assert any("no stages" in e.lower() for e in p.validate())
 class TestPipelineRegistration:
    def test_resolve_pipeline_prefers_deploy_pipeline_key(self, clean_pipeline_registry, tmp_path):
        deploy_key = "deploy_selected_pipeline"
        model_type_key = "hf_model_type_pipeline"
-        deploy_pipe = PipelineConfig(model_type=deploy_key)
+        deploy_pipe = PipelineConfig(
-        model_type_pipe = PipelineConfig(model_type=model_type_key)
+            model_type=deploy_key,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        model_type_pipe = PipelineConfig(
            model_type=model_type_key,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(deploy_pipe)
        register_pipeline(model_type_pipe)
@@ -676,7 +727,11 @@ class TestPipelineRegistration:
        tmp_path,
    ):
        model_type_key = "registered_model_type_pipeline"
-        register_pipeline(PipelineConfig(model_type=model_type_key))
+        pipeline_cfg = PipelineConfig(
            model_type=model_type_key,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipeline_cfg)
        deploy_path = tmp_path / "deploy.yaml"
        deploy_path.write_text("pipeline: missing_pipeline\n", encoding="utf-8")
@@ -711,6 +766,7 @@ class TestPipelineRegistration:
        pipe_cfg = PipelineConfig(
            model_type=pipeline_key,
            hf_architectures=("ArchitectureFallbackForTest",),
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_cfg)
@@ -741,15 +797,18 @@ class TestPipelineRegistration:
            model_type="rejected_by_predicate",
            hf_architectures=(shared_arch,),
            hf_config_predicate=rejecting_predicate,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        raises_cfg = PipelineConfig(
            model_type="raises_in_predicate",
            hf_architectures=(shared_arch,),
            hf_config_predicate=raising_predicate,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        accept_cfg = PipelineConfig(
            model_type="accepted_by_predicate",
            hf_architectures=(shared_arch,),
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(reject_cfg)
        register_pipeline(raises_cfg)
@@ -783,6 +842,7 @@ class TestPipelineRegistration:
            model_type="predicate_without_arch_match",
            hf_architectures=("DifferentArchitectureForTest",),
            hf_config_predicate=predicate,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_cfg)
@@ -806,6 +866,7 @@ class TestPipelineRegistration:
        resolved_cfg = PipelineConfig(
            model_type="callable_resolved_pipeline",
            hf_architectures=("CallableArchitectureForTest",),
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        seen_hf_configs = []
@@ -879,7 +940,10 @@ class TestPipelineRegistration:
    def test_pipeline_registration(self, clean_pipeline_registry):
        """Ensure that we can register and create a custom pipeline config."""
        new_model_type = "new_model_type"
-        pipe_cfg = PipelineConfig(model_type=new_model_type)
+        pipe_cfg = PipelineConfig(
            model_type=new_model_type,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        # Register the new PipelineConfig
        assert new_model_type not in OMNI_PIPELINES
@@ -922,7 +986,10 @@ class TestPipelineRegistration:
        def custom_resolver(
            hf_config: FakeConfig,
        ) -> PipelineConfig:
-            return PipelineConfig(model_type=resolved_type)
+            return PipelineConfig(
                model_type=resolved_type,
                stages=SINGLE_STAGE_PIPE_CFG,
            )
        # Register the new PipelineConfig
        assert new_model_type not in OMNI_PIPELINES
@@ -979,7 +1046,7 @@ class TestPipelineRegistration:
        pipe_cfg = PipelineConfig(
            model_type=pipeline_key,
            model_arch="DeployOnlyArch",
-            stages=(StagePipelineConfig(stage_id=0, model_stage="diffusion"),),
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_cfg)
@@ -1012,7 +1079,11 @@ class TestPipelineRegistration:
    def test_structured_path_loads_explicit_deploy_config_once(self, clean_pipeline_registry, tmp_path):
        pipeline_key = "single_load_pipeline"
-        register_pipeline(PipelineConfig(model_type=pipeline_key))
+        pipeline_cfg = PipelineConfig(
            model_type=pipeline_key,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipeline_cfg)
        deploy_path = tmp_path / "single_load.yaml"
        deploy_path.write_text(f"pipeline: {pipeline_key}\n", encoding="utf-8")
@@ -1042,6 +1113,7 @@ class TestPipelineRegistration:
        pipeline = PipelineConfig(
            model_type="pipeline_with_default",
            default_deploy_config_name=default_name,
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        with patch(
@@ -1061,8 +1133,16 @@ class TestPipelineRegistration:
            OmniServingCapability.COMPLETIONS,
            "pipeline_a blocks completions",
        )
-        pipe_a = PipelineConfig(model_type="detect_type", endpoint_restrictions=(restriction,))
+        pipe_a = PipelineConfig(
-        pipe_b = PipelineConfig(model_type="override_type", endpoint_restrictions=())
+            model_type="detect_type",
            endpoint_restrictions=(restriction,),
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        pipe_b = PipelineConfig(
            model_type="override_type",
            endpoint_restrictions=(),
            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_a)
        register_pipeline(pipe_b)
@@ -1144,7 +1224,7 @@ class TestDeployConfigLoading:
        stages = merge_pipeline_deploy(pipeline, deploy)
        assert deploy.session_mode == "duplex"
-        assert deploy.active_stream_window == 1
+        assert deploy.active_stream_window == max_sessions
        assert deploy.duplex_session.max_sessions == max_sessions
        assert [stage.session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
        assert [stage.to_omegaconf().session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
@@ -1324,6 +1404,13 @@ stages:
            ("step_audio2_ci.yaml", "step_audio_2_asr", 1, "text", "step_audio_2_asr"),
            ("hunyuan_video_15.yaml", "hunyuan_video_15", 1, "video", None),
            ("wan2_2_ti2v.yaml", "wan2_2_ti2v", 1, "video", None),
            (
                "minimax_h3_disaggregated.yaml",
                "minimax_h3_disaggregated",
                2,
                "video",
                "minimax_h3_disaggregated",
            ),
        ],
    )
    def test_load_new_registry_backed_deploy_configs(
@@ -1344,9 +1431,91 @@ stages:
        assert len(stages) == stage_count
        assert stages[-1].final_output is True
        assert stages[-1].final_output_type == final_output_type
        if pipeline_name == "minimax_h3_disaggregated":
            assert stages[0].yaml_engine_args["model_path_resolver"].endswith(".resolve_minimax_h3_model_root")
            assert (
                stages[1].yaml_engine_args["model_path_resolver"].endswith(".resolve_minimax_h3_diffusion_model_path")
            )
            assert "model_path_resolver" not in stages[0].yaml_extras
            assert "model_path_resolver" not in stages[1].yaml_extras
        if deploy.trust_remote_code is not None:
            assert {s.yaml_engine_args.get("trust_remote_code") for s in stages} == {deploy.trust_remote_code}
    def test_minimax_h3_resolvers_reach_production_engine_args(self, tmp_path: Path, monkeypatch):
        from vllm_omni.engine import stage_init_utils
        model_root = tmp_path / "MiniMax-H3"
        (model_root / "FL2VA" / "text_encoder").mkdir(parents=True)
        (model_root / "Ref2VA" / "text_encoder").mkdir(parents=True)
        monkeypatch.setattr(stage_init_utils, "resolve_worker_cls", lambda _engine_args: None)
        deploy = load_deploy_config(get_deploy_config_path("minimax_h3_disaggregated.yaml"))
        assert deploy.stages[1].engine_extras["model_loaded"] == {"text_encoder": False}
        stages = merge_pipeline_deploy(OMNI_PIPELINES["minimax_h3_disaggregated"], deploy)
        resolved = [stage_init_utils.build_engine_args_dict(stage.to_omegaconf(), str(model_root)) for stage in stages]
        assert resolved[0]["model"] == str(model_root / "FL2VA" / "text_encoder")
        assert resolved[1]["model"] == str(model_root / "FL2VA")
        assert all("model_path_resolver" not in engine_args for engine_args in resolved)
    def test_minimax_h3_text_encoder_tp_reaches_only_stage_zero(self):
        pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
        stages, _ = StageConfigFactory._create_legacy_from_registry(
            pipeline,
            {"text_encoder_tp_size": 4},
        )
        assert stages[0].runtime_overrides["tensor_parallel_size"] == 4
        assert "text_encoder_tp_size" not in stages[0].runtime_overrides
        assert "text_encoder_tp_size" not in stages[1].runtime_overrides
        assert stages[1].yaml_engine_args["parallel_config"]["tensor_parallel_size"] == 1
    def test_minimax_h3_disaggregated_defaults_match_validated_topology(self):
        pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
        deploy = load_deploy_config(Path(get_deploy_config_path("minimax_h3_disaggregated.yaml")))
        stages = merge_pipeline_deploy(pipeline, deploy)
        assert stages[0].yaml_engine_args["max_num_seqs"] == 1
        assert stages[0].yaml_engine_args["model_path_resolver"].endswith(".resolve_minimax_h3_model_root")
        assert stages[1].yaml_engine_args["model_path_resolver"].endswith(".resolve_minimax_h3_diffusion_model_path")
        parallel = stages[1].yaml_engine_args["parallel_config"]
        assert parallel["tensor_parallel_size"] == 1
        assert parallel["ulysses_degree"] == 4
        assert parallel["vae_patch_parallel_size"] == 4
        assert stages[1].yaml_engine_args["inline_diffusion"] is True
        assert stages[1].yaml_extras["default_sampling_params"]["num_inference_steps"] == 50
        turbo = load_deploy_config(Path(get_deploy_config_path("minimax_h3_disaggregated_turbo.yaml")))
        turbo_stages = merge_pipeline_deploy(pipeline, turbo)
        turbo_sampling = turbo_stages[1].yaml_extras["default_sampling_params"]
        assert turbo_sampling["num_inference_steps"] == 5
        assert turbo_sampling["extra_args"] == {"flow_shift": 6.0, "audio_flow_shift": 3.0}
    def test_minimax_h3_text_encoder_tp_alias_targets_stage_zero(self):
        pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
        assert normalize_pipeline_cli_overrides(pipeline, {}) == {}
        assert normalize_pipeline_cli_overrides(pipeline, {"text_encoder_tp_size": 4}) == {
            "stage_0_tensor_parallel_size": 4
        }
    def test_minimax_h3_stage_zero_tp_override_wins_alias_conflict(self):
        pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
        with pytest.warns(UserWarning, match="stage_0_tensor_parallel_size=2 takes precedence"):
            normalized = normalize_pipeline_cli_overrides(
                pipeline,
                {"text_encoder_tp_size": 4, "stage_0_tensor_parallel_size": 2},
            )
        assert normalized == {"stage_0_tensor_parallel_size": 2}
    def test_minimax_h3_rejects_text_encoder_tp_on_diffusion_stage(self):
        pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
        with pytest.raises(ValueError, match="stage_1_text_encoder_tp_size cannot be set"):
            normalize_pipeline_cli_overrides(pipeline, {"stage_1_text_encoder_tp_size": 4})
    @pytest.mark.parametrize(
        ("config_json", "model_index", "expected_pipeline"),
        [
@@ -1382,6 +1551,29 @@ stages:
        assert pipeline is not None
        assert pipeline.model_type == expected_pipeline
    def test_minimax_h3_disaggregation_is_not_auto_discovered(self):
        def get_model_file(filename, _model, revision=None):
            del revision
            if filename == "config.json":
                return None
            if filename == "model_index.json":
                return {"_class_name": "MiniMaxH3Pipeline"}
            return None
        with (
            patch.object(StageConfigFactory, "get_hf_config", return_value=None),
            patch(
                "vllm_omni.config.config_factory.get_hf_file_to_dict",
                side_effect=get_model_file,
            ),
        ):
            pipeline = StageConfigFactory.get_pipeline_config(
                model="/models/unrelated-checkpoint-name",
                trust_remote_code=False,
            )
        assert pipeline is None
    @pytest.mark.parametrize("deploy_name", ["step_audio_2.yaml", "step_audio_2_async_chunk.yaml"])
    def test_step_audio2_deploy_configs_fit_two_gpus(self, deploy_name):
        deploy = load_deploy_config(Path(get_deploy_config_path(deploy_name)))
@@ -1490,6 +1682,7 @@ stages:
                    model_stage="ar",
                    execution_type=StageExecutionType.LLM_AR,
                    requires_multimodal_data=True,
                    final_output=True,
                ),
            ),
        )
@@ -1510,6 +1703,7 @@ stages:
                    model_stage="ar",
                    execution_type=StageExecutionType.LLM_AR,
                    scheduler_cls=scheduler_cls,
                    final_output=True,
                ),
            ),
        )
@@ -1755,7 +1949,6 @@ class TestQwen3OmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "Qwen3OmniMoeForConditionalGeneration"
        assert len(p.stages) == 3
        assert p.validate() == []
    def test_thinker(self):
        p = resolve_pipeline_config(
@@ -1809,7 +2002,6 @@ class TestQwen2_5OmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "Qwen2_5OmniForConditionalGeneration"
        assert len(p.stages) == 3
        assert p.validate() == []
    def test_thinker(self):
        p = resolve_pipeline_config("qwen2_5_omni")
@@ -1858,7 +2050,6 @@ class TestQwen3TTSPipeline:
        assert p is not None
        assert p.model_arch == "Qwen3TTSTalkerForConditionalGeneration"
        assert len(p.stages) == 2
        assert p.validate() == []
    def test_talker_stage(self):
        p = resolve_pipeline_config("qwen3_tts")
@@ -1938,7 +2129,6 @@ class TestMingFlashOmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "MingFlashOmniForConditionalGeneration"
        assert len(p.stages) == 2
        assert p.validate() == []
    def test_thinker_stage(self):
        p = resolve_pipeline_config("ming_flash_omni")
@@ -1995,7 +2185,6 @@ class TestMingFlashOmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "MingFlashOmniTalkerForConditionalGeneration"
        assert len(p.stages) == 1
        assert p.validate() == []
    def test_tts_stage(self):
        p = resolve_pipeline_config("ming_flash_omni_tts")
@@ -2052,7 +2241,6 @@ class TestMingFlashOmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "MingFlashOmniForConditionalGeneration"
        assert len(p.stages) == 1
        assert p.validate() == []
    def test_thinker_only_stage(self):
        p = resolve_pipeline_config("ming_flash_omni_thinker_only")
@@ -2091,7 +2279,6 @@ class TestMingFlashOmniPipeline:
        assert p is not None
        assert p.model_arch == "MingFlashOmniForConditionalGeneration"
        assert len(p.stages) == 2
        assert p.validate() == []
    def test_image_thinker_stage(self):
        s = resolve_pipeline_config("ming_flash_omni_image").get_stage(0)
@@ -2272,6 +2459,20 @@ class TestPlatformOverrides:
        assert rocm.stages[0].enforce_eager is None
        assert rocm.stages[1].enforce_eager is True
    def test_higgs_audio_v3_rocm_uses_triton_attention(self):
        deploy_path = Path(get_deploy_config_path("higgs_multimodal_qwen3.yaml"))
        base = load_deploy_config(deploy_path)
        assert base.stages[0].engine_extras["attention_backend"] == "FLASHINFER"
        rocm = _apply_platform_overrides(base, platform="rocm")
        assert rocm.stages[0].engine_extras["attention_backend"] == "TRITON_ATTN"
        pipeline = resolve_pipeline_config("higgs_multimodal_qwen3")
        assert isinstance(pipeline, PipelineConfig)
        stages = merge_pipeline_deploy(pipeline, rocm)
        assert stages[0].yaml_engine_args["attention_backend"] == "TRITON_ATTN"
    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(
@@ -45,6 +45,7 @@ from vllm_omni.config.stage_config import (
    PipelineConfig,
    StageDeployConfig,
    StageExecutionType,
    StagePipelineConfig,
    load_deploy_config,
    merge_pipeline_deploy,
 )
@@ -92,10 +93,23 @@ def _from_pipeline_key(
    )
 def test_minimax_h3_text_encoder_tp_targets_only_structured_stage_zero() -> None:
    config = _from_pipeline_key(
        "minimax_h3_disaggregated",
        cli_overrides={"text_encoder_tp_size": 4},
    )
    assert config.stage_by_id(0).parallel_config.tensor_parallel_size == 4
    diffusion = config.stage_by_id(1)
    assert diffusion.parallel_config.tensor_parallel_size == 1
    assert diffusion.parallel_config.text_encoder_tp_size == 1
 def test_duplex_session_capacity_propagates_to_every_structured_stage() -> None:
    omni_config = _from_pipeline_key("personaplex")
    assert [stage.model_config.duplex_max_sessions for stage in omni_config.stage_configs] == [2, 2]
    assert [stage.model_config.session_mode for stage in omni_config.stage_configs] == ["duplex", "duplex"]
 def test_non_duplex_deploy_keeps_model_session_capacity_at_one(tmp_path: Path) -> None:
@@ -105,6 +119,7 @@ def test_non_duplex_deploy_keeps_model_session_capacity_at_one(tmp_path: Path) -
    omni_config = _from_pipeline_key("personaplex", deploy_config_path=str(deploy_path))
    assert [stage.model_config.duplex_max_sessions for stage in omni_config.stage_configs] == [1, 1]
    assert [stage.model_config.session_mode for stage in omni_config.stage_configs] == ["turn", "turn"]
@pytest.mark.parametrize("model_type", sorted(OMNI_PIPELINES))
@@ -132,6 +147,7 @@ def test_vllm_omni_config_from_pipeline_config_matches_merge_pipeline_deploy(mod
        engine_args = legacy_stage.yaml_engine_args
        assert omni_stage.model_config.duplex_max_sessions == engine_args.get("duplex_max_sessions", 1)
        assert omni_stage.model_config.session_mode == engine_args.get("session_mode", "turn")
        assert omni_stage.model_config.enforce_eager == engine_args.get("enforce_eager", False)
        assert omni_stage.load_config.load_format == engine_args.get("load_format", "auto")
        assert omni_stage.load_config.tokenizer_mode == engine_args.get("tokenizer_mode", "auto")
@@ -579,6 +595,7 @@ def test_sub_config_fields_match_structured_scopes():
        "interleave_mm_strings",
        "media_io_kwargs",
        "active_stream_window",
        "session_mode",
        "duplex_max_sessions",
        "enable_sleep_mode",
        "default_sampling_params",
@@ -637,10 +654,11 @@ def test_sub_config_fields_match_structured_scopes():
        "ring_degree",
        "allgather_degree",
        "ulysses_mode",
        "ulysses_a2a_permute",
        "cfg_parallel_size",
        "vae_patch_parallel_size",
        "text_encoder_tp_size",
        "vae_parallel_mode",
        "text_encoder_tp_size",
        "use_hsdp",
        "mask_sp_padding",
        "hsdp_shard_size",
@@ -842,6 +860,23 @@ def test_from_pipeline_config_derives_sequence_parallel_size_from_allgather_degr
    assert stage.parallel_config.world_size == 2
 def test_from_pipeline_config_preserves_deployed_ulysses_a2a_permute() -> None:
    pipeline = _resolve_pipeline_or_skip("dreamzero")
    deploy = DeployConfig(
        stages=[
            StageDeployConfig(
                stage_id=0,
                ulysses_a2a_permute=True,
            )
        ]
    )
    stage = VllmOmniConfig.from_pipeline_config(pipeline, user_deploy_config=deploy).stage_by_id(0)
    assert isinstance(stage, VllmOmniDiffusionStageConfig)
    assert stage.parallel_config.ulysses_a2a_permute is True
 def test_diffusion_parallel_config_accepts_four_way_guidance_parallelism():
    cfg = OmniStageDiffusionParallelConfig(cfg_parallel_size=4)
@@ -936,6 +971,35 @@ def test_structured_diffusion_config_rejects_non_boolean_compile_dynamic():
        omni_config_module._DiffusionConfigProjection(diffusion_compile_dynamic="false")
 def test_from_pipeline_config_routes_ltx2_conv_vae_extra(tmp_path):
    deploy_path = tmp_path / "dreamzero_ltx2_extras.yaml"
    deploy_path.write_text(
        "\n".join(
            [
                "pipeline: dreamzero",
                "async_chunk: false",
                "stages:",
                "  - stage_id: 0",
                "    extras:",
                "      ltx2_use_conv_vae: true",
            ]
        )
    )
    stage = _from_pipeline_key("dreamzero", deploy_config_path=str(deploy_path)).stage_by_id(0)
    assert stage.diffusion_config.extras["ltx2_use_conv_vae"] is True
 def test_stage_override_routes_ltx2_conv_vae_extra():
    stage = _from_pipeline_key(
        "dreamzero",
        cli_overrides={"stage_0_extras": {"ltx2_use_conv_vae": True}},
    ).stage_by_id(0)
    assert stage.diffusion_config.extras["ltx2_use_conv_vae"] is True
 def test_structured_diffusion_config_rejects_invalid_compile_granularity():
    with pytest.raises(ValidationError, match="diffusion_compile_granularity"):
        omni_config_module._DiffusionConfigProjection(diffusion_compile_granularity="block")
@@ -976,7 +1040,10 @@ def test_from_pipeline_config_uses_hf_config_for_callable_resolver():
 def test_from_pipeline_config_accepts_pre_resolved_pipeline():
-    resolved_pipeline = PipelineConfig(model_type="callable_resolved_variant")
+    resolved_pipeline = PipelineConfig(
        model_type="callable_resolved_variant",
        stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),),
    )
    omni_config = VllmOmniConfig.from_pipeline_config(resolved_pipeline)
@@ -1009,6 +1076,7 @@ def test_from_pipeline_config_default_deploy_name_ignores_cwd(monkeypatch, tmp_p
    pipeline = PipelineConfig(
        model_type="pipeline_with_default",
        default_deploy_config_name=default_name,
        stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),),
    )
    loaded_paths = []
@@ -1,16 +1,22 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for out of tree registration to OMNI_PIPELINES."""
 import pytest
 from transformers import PretrainedConfig
 from vllm_omni.config.pipeline_registry import OMNI_PIPELINES, register_pipeline
-from vllm_omni.config.stage_config import PipelineConfig, pipeline_cfg_resolver
+from vllm_omni.config.stage_config import PipelineConfig, StagePipelineConfig, pipeline_cfg_resolver
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 def build_fake_pipeline_config(model_type: str) -> PipelineConfig:
    return PipelineConfig(
        model_type=model_type, stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),)
    )
@pytest.fixture
 def custom_resolver():
    """Build a reusable custom resolver for PipelineConfigs."""
@@ -22,7 +28,7 @@ def custom_resolver():
    def custom_resolver(
        hf_config: CustomConfigType,
    ) -> PipelineConfig:
-        return PipelineConfig(model_type="resolved_type")
+        return build_fake_pipeline_config("resolved_type")
    return custom_resolver
@@ -30,7 +36,7 @@ def custom_resolver():
 def test_register_pipeline_config(clean_pipeline_registry):
    """Ensure that we can register a custom pipeline config to OMNI_PIPELINES."""
    new_model_type = "new_model_type"
-    pipe_cfg = PipelineConfig(model_type=new_model_type)
+    pipe_cfg = build_fake_pipeline_config(new_model_type)
    assert new_model_type not in OMNI_PIPELINES
    register_pipeline(pipe_cfg)
    assert new_model_type in OMNI_PIPELINES
@@ -41,7 +47,7 @@ def test_register_pipeline_config_with_model_type(clean_pipeline_registry):
    """Ensure that we can register a custom pipeline config with an explicit model_type to OMNI_PIPELINES."""
    new_model_type = "new_model_type"
    unused_model_type = "foo"
-    pipe_cfg = PipelineConfig(model_type=unused_model_type)
+    pipe_cfg = build_fake_pipeline_config(unused_model_type)
    assert new_model_type not in OMNI_PIPELINES
    assert unused_model_type not in OMNI_PIPELINES
@@ -66,3 +72,10 @@ def test_register_resolver_requires_model_type(custom_resolver, clean_pipeline_r
    """Ensure that registering a custom resolver to OMNI_PIPELINES requires an explicit model_type."""
    with pytest.raises(ValueError):
        register_pipeline(custom_resolver)
 def test_minimax_h3_disaggregation_is_explicit_opt_in():
    assert "minimax_h3" not in OMNI_PIPELINES
    pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
    assert isinstance(pipeline, PipelineConfig)
    assert pipeline.model_type == "minimax_h3_disaggregated"
@@ -25,9 +25,9 @@ pytest_plugins = (
    "tests.helpers.fixtures.clean",
    "tests.helpers.fixtures.log",
    "tests.helpers.fixtures.media",
-    "tests.helpers.fixtures.run_args",
+    "tests.helpers.fixtures.pytest_collection",
    "tests.helpers.fixtures.pytest_run_args",
    "tests.helpers.fixtures.runtime",
    "tests.helpers.fixtures.xdist",
 )
@@ -0,0 +1,89 @@
 """Tests for ``OmniARScheduler._drop_aborted_queued_requests``.
 ``schedule()`` sweeps ``FINISHED_ABORTED`` requests out of the queues before
 handing control to upstream, because upstream ``Scheduler.schedule()`` raises
 ``RuntimeError: Invalid request status: FINISHED_ABORTED`` for any request it
 admits in a finished state -- which kills the stage's engine core, not just
 the request.
 Regression for the sweep missing ``skipped_waiting``, the third queue
 upstream admits from. An aborted duplex session parked there was re-selected
 by ``_select_waiting_queue_for_scheduling`` on a later tick and crashed the
 stage.
 """
 from __future__ import annotations
 import pytest
 from vllm.v1.core.sched.request_queue import SchedulingPolicy, create_request_queue
 from vllm.v1.request import RequestStatus
 import vllm_omni.core.sched.omni_ar_scheduler as ar_sched_mod
 from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 _POLICIES = [
    pytest.param(SchedulingPolicy.FCFS, id="fcfs"),
    pytest.param(SchedulingPolicy.PRIORITY, id="priority"),
 ]
 class _StubRequest:
    """Minimal ``Request`` stub with the surface the sweep exercises.
    Carries the ``(priority, arrival_time)`` ordering ``PriorityRequestQueue``
    relies on so the same stub works under both scheduling policies.
    """
    def __init__(self, request_id: str, status: RequestStatus, arrival_time: float = 0.0) -> None:
        self.request_id = request_id
        self.status = status
        self.priority = 0
        self.arrival_time = arrival_time
    def __lt__(self, other: _StubRequest) -> bool:
        return (self.priority, self.arrival_time) < (other.priority, other.arrival_time)
 def _make_scheduler(policy: SchedulingPolicy, *, waiting=(), skipped_waiting=(), running=()):
    """Bare scheduler with just the surface the sweep reads."""
    scheduler = OmniARScheduler.__new__(OmniARScheduler)
    scheduler.waiting = create_request_queue(policy)
    for req in waiting:
        scheduler.waiting.add_request(req)
    scheduler.skipped_waiting = create_request_queue(policy)
    for req in skipped_waiting:
        scheduler.skipped_waiting.add_request(req)
    scheduler.running = list(running)
    return scheduler
@pytest.mark.parametrize("policy", _POLICIES)
 def test_schedule_sweeps_skipped_waiting_before_upstream_selection(
    monkeypatch: pytest.MonkeyPatch,
    policy: SchedulingPolicy,
 ) -> None:
    """The public schedule path must remove aborted skipped requests first."""
    aborted = _StubRequest("req-aborted", RequestStatus.FINISHED_ABORTED)
    scheduler = _make_scheduler(policy, skipped_waiting=[aborted])
    scheduler.chunk_transfer_adapter = None
    scheduler.input_coordinator = None
    scheduler.num_waiting_for_streaming_input = 0
    scheduler._process_pending_omni_inputs = lambda model_mode: None
    scheduler._resync_streaming_input_counter = lambda: None
    scheduler._should_defer_waiting_admission = lambda: False
    scheduler._restore_omni_wait_queues = lambda: None
    scheduler._postprocess_omni_schedule_output = lambda *args, **kwargs: None
    scheduler.get_finished_requests_needing_kv_transfer = lambda: {}
    scheduler._wrap_omni_scheduler_output = lambda output, **kwargs: output
    sentinel = object()
    def upstream_schedule(_self, _throttle_prefills: bool = False):
        assert list(scheduler.skipped_waiting) == []
        return sentinel
    monkeypatch.setattr(ar_sched_mod.VLLMScheduler, "schedule", upstream_schedule)
    assert OmniARScheduler.schedule(scheduler) is sentinel
@@ -142,13 +142,16 @@ def test_exact_drain_delivers_new_segment_frame() -> None:
    session.num_computed_tokens = 6
    session.num_output_placeholders = 1
    session.num_in_flight_tokens = 1
    session.async_tokens_to_discard = 1
    _replace_streaming_session(session)
    assert session.num_stale_output_tokens == 1
    assert session.async_tokens_to_discard == 1
    sched = _make_drain_sched(session)
    assert _run_step(sched, session, num_scheduled=1, token=42) is False  # late frame dropped
    assert session.num_stale_output_tokens == 0
    assert session.async_tokens_to_discard == 0
    assert _run_step(sched, session, num_scheduled=1, token=43) is True  # new segment survives
@@ -18,17 +18,19 @@ from vllm.sampling_params import SamplingParams
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.v1.request import Request, RequestStatus, StreamingUpdate
-from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARAsyncScheduler, OmniARScheduler
 # isort: on
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-def _make_scheduler(*, stage_id: int = 0) -> OmniARScheduler:
+def _make_scheduler(*, stage_id: int = 0, session_mode: str = "turn") -> OmniARScheduler:
    sched = OmniARScheduler.__new__(OmniARScheduler)
    sched._new_prompt_len_snapshot = {}
-    sched.vllm_config = SimpleNamespace(model_config=SimpleNamespace(stage_id=stage_id))
+    sched.vllm_config = SimpleNamespace(
        model_config=SimpleNamespace(stage_id=stage_id, session_mode=session_mode),
    )
    sched.num_waiting_for_streaming_input = 0
    sched.log_stats = False
    sched.chunk_transfer_adapter = None
@@ -185,6 +187,7 @@ def test_resumable_segment_boundary_keeps_pre_transition_send_watermark() -> Non
    def replace_with_next_segment(request: Request) -> bool:
        request.num_computed_tokens = 0
        request._omni_segment_generation = 1
        request.status = RequestStatus.WAITING
        return False
@@ -201,6 +204,7 @@ def test_resumable_segment_boundary_keeps_pre_transition_send_watermark() -> Non
        True,
        new_token_ids=[42],
        confirmed_num_computed_tokens=26,
        segment_generation=0,
    )
@@ -286,6 +290,140 @@ def test_queued_streaming_update_on_async_stop_fences_in_flight_once() -> None:
    assert session.num_stale_output_tokens == session.num_in_flight_tokens == 1
 def test_stale_async_frame_is_dropped_before_output_processing() -> None:
    session = _make_request()
    session.status = RequestStatus.RUNNING
    session.num_in_flight_tokens = 2
    session.num_computed_tokens = session.num_tokens + 1
    session.num_output_placeholders = 1
    session.async_tokens_to_discard = 1
    session.sampling_params = SimpleNamespace(num_logprobs=1)
    num_computed_tokens = session.num_computed_tokens
    num_output_placeholders = session.num_output_placeholders
    sched = MagicMock()
    sched.requests = {session.request_id: session}
    sched.perf_metrics = None
    sched.structured_output_manager.should_advance.return_value = False
    def discard_stale_output(request: Request, token_ids: list[int]) -> tuple[list[int], bool]:
        request.async_tokens_to_discard = 0
        return token_ids, False
    sched._update_request_with_output.side_effect = discard_stale_output
    sched._process_kv_transfer_trigger.return_value = False
    sched.chunk_transfer_adapter = MagicMock()
    sched.running = [session]
    sched.waiting_for_transfer_free = set()
    sched.transfer_triggered_requests = set()
    sched.active_kv_transfers = set()
    sched.pending_stop_after_extraction = set()
    sched.connector = None
    sched.kv_cache_manager.take_events.return_value = None
    sched.kv_cache_manager.estimate_cached_tokens.return_value = 0
    sched.finished_req_ids_dict = {}
    sched.make_stats.return_value = None
    scheduler_output = MagicMock(spec=SchedulerOutput)
    scheduler_output.num_scheduled_tokens = {session.request_id: 1}
    scheduler_output.scheduled_spec_decode_tokens = {}
    scheduler_output.num_invalid_spec_tokens = 0
    model_runner_output = MagicMock(spec=ModelRunnerOutput)
    model_runner_output.sampled_token_ids = [[42]]
    model_runner_output.logprobs = None
    model_runner_output.prompt_logprobs_dict = {}
    model_runner_output.pooler_output = None
    model_runner_output.multimodal_outputs = None
    model_runner_output.inter_stage_outputs = [{"hidden": object()}]
    model_runner_output.num_nans_in_logits = None
    model_runner_output.kv_connector_output = None
    model_runner_output.cudagraph_stats = None
    model_runner_output.req_id_to_index = {session.request_id: 0}
    model_runner_output.routed_experts = None
    OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
    assert session.async_tokens_to_discard == 0
    assert session.status == RequestStatus.RUNNING
    assert session.num_computed_tokens == num_computed_tokens
    assert session.num_output_placeholders == num_output_placeholders
    sched.chunk_transfer_adapter.save_async.assert_not_called()
    session.sampling_params = SamplingParams(max_tokens=8)
    next_payload = {"hidden": object()}
    sched._update_request_with_output.side_effect = None
    sched._update_request_with_output.return_value = ([43], False)
    model_runner_output.sampled_token_ids = [[43]]
    model_runner_output.inter_stage_outputs = [next_payload]
    OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
    sched.chunk_transfer_adapter.save_async.assert_called_once_with(
        next_payload,
        session,
        False,
        new_token_ids=[43],
        confirmed_num_computed_tokens=None,
    )
 def test_legacy_stale_async_marker_bypasses_real_placeholder_accounting() -> None:
    request = _make_request()
    request.status = RequestStatus.RUNNING
    request.num_in_flight_tokens = 1
    request.num_output_placeholders = 0
    request.async_tokens_to_discard = 1
    request.num_stale_output_tokens = 0
    sched = OmniARAsyncScheduler.__new__(OmniARAsyncScheduler)
    sched.requests = {request.request_id: request}
    sched.perf_metrics = None
    sched.chunk_transfer_adapter = None
    sched.connector = None
    sched.ec_connector = None
    sched.waiting_for_transfer_free = set()
    sched.transfer_triggered_requests = set()
    sched.active_kv_transfers = set()
    sched.pending_stop_after_extraction = set()
    sched.finished_req_ids_dict = {}
    sched._new_prompt_len_snapshot = {}
    sched.kv_cache_manager = MagicMock()
    sched.kv_cache_manager.take_events.return_value = None
    sched._remove_stopped_requests_from_queues = MagicMock()
    sched._handle_failed_kv_load_outputs = MagicMock(return_value=[])
    sched._cleanup_kv_tracking = MagicMock()
    sched._aggregate_kv_connector_stats = MagicMock(return_value=None)
    sched._publish_kv_cache_events = MagicMock()
    sched._attach_finished_request_sets = MagicMock()
    sched._attach_scheduler_stats = MagicMock()
    sched._capture_omni_connector_output = MagicMock()
    scheduler_output = MagicMock(spec=SchedulerOutput)
    scheduler_output.num_scheduled_tokens = {request.request_id: 1}
    scheduler_output.scheduled_spec_decode_tokens = {}
    scheduler_output.num_invalid_spec_tokens = 0
    model_runner_output = MagicMock(spec=ModelRunnerOutput)
    model_runner_output.sampled_token_ids = [[42]]
    model_runner_output.logprobs = None
    model_runner_output.prompt_logprobs_dict = {}
    model_runner_output.pooler_output = None
    model_runner_output.multimodal_outputs = None
    model_runner_output.inter_stage_outputs = None
    model_runner_output.num_nans_in_logits = None
    model_runner_output.kv_connector_output = None
    model_runner_output.cudagraph_stats = None
    model_runner_output.req_id_to_index = {request.request_id: 0}
    model_runner_output.routed_experts = None
    OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
    assert request.async_tokens_to_discard == 0
    assert request.num_output_placeholders == 0
    assert request._output_token_ids == []
 def test_stage0_streaming_update_discards_outstanding_async_placeholder_token() -> None:
    sched = _make_scheduler(stage_id=0)
    session = _make_request()
@@ -446,11 +584,15 @@ def test_chunk_segment_cleanup_keeps_requeued_resumable_receiver() -> None:
        return result
    sched = OmniARScheduler.__new__(OmniARScheduler)
    sched.vllm_config = SimpleNamespace(model_config=SimpleNamespace(session_mode="duplex"))
    sched.running = []
    sched.waiting = queue()
    sched.skipped_waiting = queue(session)
    sched.num_waiting_for_streaming_input = 1
-    sched.chunk_transfer_adapter = SimpleNamespace(segment_finished_requests={session.request_id})
+    sched.chunk_transfer_adapter = SimpleNamespace(
        receives_chunks=True,
        segment_finished_requests={session.request_id},
    )
    sched._resume_downstream_chunk_receiver(session)
    sched._remove_stopped_requests_from_queues(set(), {session})
@@ -460,3 +602,37 @@ def test_chunk_segment_cleanup_keeps_requeued_resumable_receiver() -> None:
    assert session not in sched.skipped_waiting.requests
    assert sched.num_waiting_for_streaming_input == 0
    assert session.request_id not in sched.chunk_transfer_adapter.segment_finished_requests
@pytest.mark.parametrize(
    ("receives_chunks", "session_mode"),
    [
        (False, "duplex"),
        (True, "turn"),
    ],
 )
 def test_chunk_segment_cleanup_keeps_explicit_update_stage_parked(
    receives_chunks: bool,
    session_mode: str,
 ) -> None:
    """Only duplex connector-driven receivers resume without an update."""
    session = _make_request()
    session.status = RequestStatus.WAITING_FOR_STREAMING_REQ
    sched = OmniARScheduler.__new__(OmniARScheduler)
    sched.vllm_config = SimpleNamespace(model_config=SimpleNamespace(session_mode=session_mode))
    sched.num_waiting_for_streaming_input = 1
    sched.chunk_transfer_adapter = SimpleNamespace(
        receives_chunks=receives_chunks,
        segment_finished_requests={session.request_id},
    )
    sched.skipped_waiting = MagicMock()
    sched._enqueue_waiting_request = MagicMock()
    sched._resume_downstream_chunk_receiver(session)
    assert session.status == RequestStatus.WAITING_FOR_STREAMING_REQ
    assert sched.num_waiting_for_streaming_input == 1
    assert session.request_id not in sched.chunk_transfer_adapter.segment_finished_requests
    sched.skipped_waiting.remove_requests.assert_not_called()
    sched._enqueue_waiting_request.assert_not_called()
@@ -76,6 +76,9 @@ def _make_scheduler(
        else:
            hf_config.voxcpm2_runtime_config = runtime_config
    sched.vllm_config = SimpleNamespace(model_config=model_config)
    # Upstream ``Scheduler.__init__`` always creates ``skipped_waiting``; the
    # abort sweep at the top of ``schedule()`` reads it.
    sched.skipped_waiting = _MockQueue()
    return sched
@@ -166,3 +166,15 @@ def test_fast_path_schedule_defer_disabled_leaves_fence() -> None:
    OmniGenerationScheduler.schedule(sched)
    assert sched.sched_step_seq == 0
 def test_generation_schedule_sweeps_and_resyncs_around_pending_inputs() -> None:
    sched, request = _make_fast_path_sched(defer_block_free=False)
    calls: list[str] = []
    sched._drop_aborted_queued_requests.side_effect = lambda: calls.append("sweep")
    sched._process_pending_omni_inputs.side_effect = lambda model_mode: calls.append(model_mode)
    sched._resync_streaming_input_counter.side_effect = lambda: calls.append("resync")
    OmniGenerationScheduler.schedule(sched)
    assert calls == ["sweep", "generation", "sweep", "resync"]
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Unit coverage for the two stage-input timeout hooks.
 ``_process_pending_input_timeouts`` covers the full-payload transport;
@@ -253,6 +253,38 @@ def test_log_failed_chunk_sends_noop_without_adapter():
    scheduler._log_failed_chunk_sends()
 class _FakeReceiveFailureAdapter:
    def __init__(self, failures):
        self._failures = dict(failures)
    def collect_failed_receive_request_ids(self):
        failures, self._failures = self._failures, {}
        return failures
 class _FakeReceiveFailureScheduler(OmniSchedulerMixin):
    def __init__(self, requests, adapter):
        self.requests = requests
        self.chunk_transfer_adapter = adapter
        self.finish_calls = []
    def finish_requests(self, req_ids, status):
        self.finish_calls.append((set(req_ids), status))
 def test_process_chunk_receive_failures_delegates_to_finish_requests():
    request_id = "invalid-chunk"
    adapter = _FakeReceiveFailureAdapter({request_id: "prompt exceeds limit"})
    scheduler = _FakeReceiveFailureScheduler({request_id: SimpleNamespace(request_id=request_id)}, adapter)
    scheduler._process_chunk_receive_failures()
    assert len(scheduler.finish_calls) == 1
    finished_ids, status = scheduler.finish_calls[0]
    assert finished_ids == {request_id}
    assert getattr(status, "name", str(status)).endswith("FINISHED_ERROR")
 # --- R1.4: VLLM_OMNI_INPUT_WAIT_TIMEOUT_S is parsed at import time, so each of
 # these reloads the module under a different environment. One knob now gates the
 # deadline on both transports, which is why a silent 0 matters more than it did
@@ -0,0 +1,168 @@
 """Tests for the ``num_waiting_for_streaming_input`` resync placed in
 ``OmniARScheduler.finish_requests`` / ``OmniGenerationScheduler.finish_requests``
 and in ``OmniARScheduler.schedule``.
 Upstream keeps that counter incrementally (``+1`` when a resumable request
 parks in a waiting queue as ``WAITING_FOR_STREAMING_REQ``, ``-1`` when the
 next update arrives or the request is finished) and
 ``Scheduler.get_num_unfinished_requests`` *subtracts* it from the waiting
 queues. Omni rewrites ``request.status`` outside both hooks -- the
 chunk-transfer adapter's park/restore and
 ``_realign_request_status_to_queues`` -- so a counted request can leave
 ``WAITING_FOR_STREAMING_REQ`` without the matching decrement and inflate the
 counter for the lifetime of the process.
 An inflated counter does not degrade throughput, it hangs the stage:
 ``EngineCore.has_work()`` reads false while a live request sits in
 ``waiting``, so the engine blocks in ``input_queue.get()`` and never calls
 ``schedule()`` again. Regression for the Qwen3-Omni duplex symptom where the
 server answered the first conversation after a boot and was silent from the
 second on while still reporting healthy.
 """
 from __future__ import annotations
 import pytest
 from vllm.v1.core.sched.request_queue import SchedulingPolicy, create_request_queue
 from vllm.v1.request import RequestStatus
 import vllm_omni.core.sched.omni_ar_scheduler as ar_sched_mod
 import vllm_omni.core.sched.omni_generation_scheduler as gen_sched_mod
 from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
 from vllm_omni.core.sched.omni_generation_scheduler import OmniGenerationScheduler
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 class _StubRequest:
    def __init__(self, request_id: str, status: RequestStatus) -> None:
        self.request_id = request_id
        self.status = status
    def is_finished(self) -> bool:
        return RequestStatus.is_finished(self.status)
 _SCHEDULER_PARAMS = [
    pytest.param(OmniARScheduler, ar_sched_mod, id="ar"),
    pytest.param(OmniGenerationScheduler, gen_sched_mod, id="generation"),
 ]
 def _make_scheduler(scheduler_cls, *, requests, running, waiting, counter, skipped=None):
    scheduler = scheduler_cls.__new__(scheduler_cls)
    scheduler.chunk_transfer_adapter = None
    scheduler.input_coordinator = None
    scheduler.requests = requests
    scheduler.running = running
    scheduler.waiting = waiting
    scheduler.skipped_waiting = skipped if skipped is not None else []
    scheduler.num_waiting_for_streaming_input = counter
    return scheduler
@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
 def test_finish_requests_clears_counter_left_by_a_stomped_status(
    monkeypatch: pytest.MonkeyPatch,
    scheduler_cls,
    scheduler_mod,
 ) -> None:
    """The abort case that wedged the talker.
    The request was counted while parked for streaming input, then had its
    status rewritten to ``RUNNING`` by omni's queue surgery, so upstream's
    ``finish_requests`` took the running branch and never decremented. With
    no parked requests left, the counter must land at 0.
    """
    aborted = _StubRequest("req-aborted", RequestStatus.RUNNING)
    scheduler = _make_scheduler(
        scheduler_cls,
        requests={},  # upstream already dropped it
        running=[],
        waiting=[],
        counter=1,
    )
    monkeypatch.setattr(
        scheduler_mod.VLLMScheduler,
        "finish_requests",
        lambda self, request_ids, finished_status: [aborted],
    )
    scheduler_cls.finish_requests(scheduler, [aborted.request_id], RequestStatus.FINISHED_ABORTED)
    assert scheduler.num_waiting_for_streaming_input == 0
 def test_resync_counts_parked_requests_in_skipped_waiting() -> None:
    """``get_num_unfinished_requests`` sums ``waiting`` and ``skipped_waiting``,
    so the resync has to look at both or it under-counts and reintroduces the
    spin the counter exists to prevent."""
    parked = _StubRequest("req-skipped", RequestStatus.WAITING_FOR_STREAMING_REQ)
    scheduler = _make_scheduler(
        OmniARScheduler,
        requests={parked.request_id: parked},
        running=[],
        waiting=[],
        counter=0,
        skipped=[parked],
    )
    scheduler._resync_streaming_input_counter()
    assert scheduler.num_waiting_for_streaming_input == 1
 def test_ar_schedule_resyncs_before_delegating_upstream(monkeypatch: pytest.MonkeyPatch) -> None:
    """Pins placement in ``schedule``: the chunk-transfer adapter rewrites
    status during ``process_pending_chunks``, so the resync has to run after
    that and before upstream schedules, while ``self.waiting`` is still the
    real queue."""
    parked = _StubRequest("req-resumed", RequestStatus.RUNNING)
    scheduler = OmniARScheduler.__new__(OmniARScheduler)
    scheduler.chunk_transfer_adapter = None
    scheduler.input_coordinator = None
    scheduler.requests = {parked.request_id: parked}
    scheduler.running = [parked]
    scheduler.waiting = []
    scheduler.skipped_waiting = []
    scheduler.num_waiting_for_streaming_input = 1  # leaked by the status rewrite
    scheduler.policy = "fcfs"
    seen: dict[str, int] = {}
    monkeypatch.setattr(
        ar_sched_mod.OmniSchedulerMixin, "_consume_pending_connector_output", lambda self, model_mode: None
    )
    monkeypatch.setattr(ar_sched_mod.OmniSchedulerMixin, "_process_pending_input_timeouts", lambda self: None)
    monkeypatch.setattr(OmniARScheduler, "_should_defer_waiting_admission", lambda self: False)
    def fake_schedule(self, throttle_prefills=False):
        seen["counter"] = self.num_waiting_for_streaming_input
        raise _StopScheduleError
    class _StopScheduleError(Exception):
        pass
    monkeypatch.setattr(ar_sched_mod.VLLMScheduler, "schedule", fake_schedule)
    with pytest.raises(_StopScheduleError):
        OmniARScheduler.schedule(scheduler)
    assert seen["counter"] == 0
 def test_generation_scheduler_sweep_removes_aborted_requests_from_all_queues() -> None:
    aborted = _StubRequest("req-generation-aborted", RequestStatus.FINISHED_ABORTED)
    scheduler = OmniGenerationScheduler.__new__(OmniGenerationScheduler)
    scheduler.waiting = create_request_queue(SchedulingPolicy.FCFS)
    scheduler.waiting.add_request(aborted)
    scheduler.skipped_waiting = create_request_queue(SchedulingPolicy.FCFS)
    scheduler.skipped_waiting.add_request(aborted)
    scheduler.running = [aborted]
    scheduler._drop_aborted_queued_requests()
    assert list(scheduler.waiting) == []
    assert list(scheduler.skipped_waiting) == []
    assert scheduler.running == []
@@ -1,4 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 import json
 import math
 import os
 import threading
 from collections.abc import Callable
@@ -159,6 +163,18 @@ def benchmark_params(request):
 def assert_result(result, params, num_prompt) -> None:
    assert result["completed"] == num_prompt, "Request failures exist"
    baseline = params.get("baseline")
    hardware = result.get("Hardware")
    hardware_baseline = baseline.get(hardware) if isinstance(baseline, dict) and isinstance(hardware, str) else None
    if isinstance(hardware_baseline, dict) and "mean_tpot_ms" in hardware_baseline:
        num_tpot_samples = result.get("num_tpot_samples")
        mean_tpot_ms = result.get("mean_tpot_ms")
        assert isinstance(num_tpot_samples, int) and not isinstance(num_tpot_samples, bool) and num_tpot_samples > 0, (
            "TPOT baseline is configured, but no measurable TPOT samples were produced"
        )
        assert (
            isinstance(mean_tpot_ms, int | float) and not isinstance(mean_tpot_ms, bool) and math.isfinite(mean_tpot_ms)
        ), "TPOT baseline is configured, but mean_tpot_ms is not finite"
    expected_audio_turns = params.get("expected_duplex_audio_turns_per_session")
    if expected_audio_turns is not None:
        session_metrics = result.get("duplex_session_metrics")
@@ -0,0 +1,106 @@
 [
  {
    "test_name": "test_hunyuanvideo15_i2v_single_device",
    "mark": [
      {
        "hardware_marks": {
          "res": {
            "cuda": "H100"
          },
          "num_cards": 1
        }
      },
      "full_model",
      "diffusion"
    ],
    "description": "Single-device baseline",
    "server_type": "vllm-omni",
    "server_params": {
      "model": "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v",
      "serve_args": {
        "flow-shift": 5.0,
        "enable-diffusion-pipeline-profiler": true
      }
    },
    "benchmark_params": [
      {
        "name": "832x480_frames33_steps4",
        "dataset": "random",
        "task": "i2v",
        "num-prompts": 10,
        "max-concurrency": 1,
        "num-input-images": 1,
        "seed": 42,
        "enable-negative-prompt": true,
        "random-request-config": [
          {
            "width": 832,
            "height": 480,
            "num_inference_steps": 4,
            "num_frames": 33,
            "fps": 24,
            "weight": 1
          }
        ],
        "extra-body": {
          "guidance_scale": 6.0,
          "flow_shift": 5.0
        }
      }
    ]
  },
  {
    "test_name": "test_hunyuanvideo15_i2v_cachedit_tp2_vae_patch2",
    "mark": [
      {
        "hardware_marks": {
          "res": {
            "cuda": "H100"
          },
          "num_cards": 2
        }
      },
      "full_model",
      "diffusion"
    ],
    "description": "CacheDiT + TP=2 + VAE patch parallel=2 + VAE tiling",
    "server_type": "vllm-omni",
    "server_params": {
      "model": "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v",
      "serve_args": {
        "flow-shift": 5.0,
        "cache-backend": "cache_dit",
        "tensor-parallel-size": 2,
        "vae-patch-parallel-size": 2,
        "vae-use-tiling": true,
        "enable-diffusion-pipeline-profiler": true
      }
    },
    "benchmark_params": [
      {
        "name": "832x480_frames33_steps4",
        "dataset": "random",
        "task": "i2v",
        "num-prompts": 10,
        "max-concurrency": 1,
        "num-input-images": 1,
        "seed": 42,
        "enable-negative-prompt": true,
        "random-request-config": [
          {
            "width": 832,
            "height": 480,
            "num_inference_steps": 4,
            "num_frames": 33,
            "fps": 24,
            "weight": 1
          }
        ],
        "extra-body": {
          "guidance_scale": 6.0,
          "flow_shift": 5.0
        }
      }
    ]
  }
 ]
@@ -0,0 +1,104 @@
 [
  {
    "test_name": "test_hunyuanvideo15_t2v_single_device",
    "mark": [
      {
        "hardware_marks": {
          "res": {
            "cuda": "H100"
          },
          "num_cards": 1
        }
      },
      "full_model",
      "diffusion"
    ],
    "description": "Single-device baseline",
    "server_type": "vllm-omni",
    "server_params": {
      "model": "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v",
      "serve_args": {
        "flow-shift": 5.0,
        "enable-diffusion-pipeline-profiler": true
      }
    },
    "benchmark_params": [
      {
        "name": "832x480_frames33_steps4",
        "dataset": "random",
        "task": "t2v",
        "num-prompts": 10,
        "max-concurrency": 1,
        "seed": 42,
        "enable-negative-prompt": true,
        "random-request-config": [
          {
            "width": 832,
            "height": 480,
            "num_inference_steps": 4,
            "num_frames": 33,
            "fps": 24,
            "weight": 1
          }
        ],
        "extra-body": {
          "guidance_scale": 6.0,
          "flow_shift": 5.0
        }
      }
    ]
  },
  {
    "test_name": "test_hunyuanvideo15_t2v_cachedit_tp2_vae_patch2",
    "mark": [
      {
        "hardware_marks": {
          "res": {
            "cuda": "H100"
          },
          "num_cards": 2
        }
      },
      "full_model",
      "diffusion"
    ],
    "description": "CacheDiT + TP=2 + VAE patch parallel=2 + VAE tiling",
    "server_type": "vllm-omni",
    "server_params": {
      "model": "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v",
      "serve_args": {
        "flow-shift": 5.0,
        "cache-backend": "cache_dit",
        "tensor-parallel-size": 2,
        "vae-patch-parallel-size": 2,
        "vae-use-tiling": true,
        "enable-diffusion-pipeline-profiler": true
      }
    },
    "benchmark_params": [
      {
        "name": "832x480_frames33_steps4",
        "dataset": "random",
        "task": "t2v",
        "num-prompts": 10,
        "max-concurrency": 1,
        "seed": 42,
        "enable-negative-prompt": true,
        "random-request-config": [
          {
            "width": 832,
            "height": 480,
            "num_inference_steps": 4,
            "num_frames": 33,
            "fps": 24,
            "weight": 1
          }
        ],
        "extra-body": {
          "guidance_scale": 6.0,
          "flow_shift": 5.0
        }
      }
    ]
  }
 ]
@@ -43,7 +43,7 @@
        "extra_body": {
          "save_duplex_request_metrics": true
        },
-        "percentile-metrics": "ttft,e2el,audio_ttfp,audio_rtf,audio_duration",
+        "percentile-metrics": "ttft,tpot,itl,e2el,audio_ttfp,audio_rtf,audio_duration",
        "baseline": {
          "H100": {
            "request_throughput": [
@@ -46,9 +46,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              3646.797
            ],
            "mean_audio_rtf": [
              0.1098
            ]
@@ -82,9 +79,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              4059.0878
            ],
            "mean_audio_rtf": [
              0.1335
            ]
@@ -120,9 +114,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              5639.5769
            ],
            "mean_audio_rtf": [
              0.1798
            ]
@@ -45,13 +45,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              15793.1209,
              30658.6632,
              37463.4943,
              47990.5839,
              64980.2199
            ],
            "mean_audio_rtf": [
              0.169,
              0.2041,
@@ -87,9 +80,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              4556.4347
            ],
            "mean_audio_rtf": [
              0.138
            ]
@@ -123,9 +113,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              4768.0972
            ],
            "mean_audio_rtf": [
              0.1609
            ]
@@ -161,9 +148,6 @@
        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
        "baseline": {
          "H100": {
            "mean_e2el_ms": [
              6395.3517
            ],
            "mean_audio_rtf": [
              0.2048
            ]
@@ -1,3 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for DFX runner metadata field exclusion."""
 import json
@@ -105,7 +108,7 @@ def test_resolve_pytest_marks_hardware_dict_with_extra():
    assert "H100" in names
    assert "cuda" in names
    assert "gpu" in names
-    assert "distributed_cuda" in names
+    assert "cards_2" in names
    assert "full_model" in names
    assert "diffusion" in names
    assert "local_model" in names
@@ -512,3 +515,41 @@ def test_omni_duplex_expected_audio_turns_rejects_incomplete_session():
            {"expected_duplex_audio_turns_per_session": 4},
            1,
        )
 def test_omni_tpot_baseline_accepts_measured_finite_sample():
    from tests.dfx.perf.scripts.run_benchmark import assert_result
    assert_result(
        {
            "completed": 1,
            "Hardware": "H100",
            "num_tpot_samples": 1,
            "mean_tpot_ms": 10.0,
        },
        {"baseline": {"H100": {"mean_tpot_ms": 20.0}}},
        1,
    )
@pytest.mark.parametrize(
    ("num_tpot_samples", "mean_tpot_ms", "match"),
    [
        (0, float("nan"), "no measurable TPOT samples"),
        (1, float("nan"), "mean_tpot_ms is not finite"),
    ],
 )
 def test_omni_tpot_baseline_rejects_missing_or_nonfinite_sample(num_tpot_samples, mean_tpot_ms, match):
    from tests.dfx.perf.scripts.run_benchmark import assert_result
    with pytest.raises(AssertionError, match=match):
        assert_result(
            {
                "completed": 1,
                "Hardware": "H100",
                "num_tpot_samples": num_tpot_samples,
                "mean_tpot_ms": mean_tpot_ms,
            },
            {"baseline": {"H100": {"mean_tpot_ms": 20.0}}},
            1,
        )
@@ -70,10 +70,14 @@ _PARAMS = [
        pytest.param({"guidance_scale": -1.0}, "guidance_scale", id="guidance_scale_negative", marks=_SKIP_ISSUE_3649),
        pytest.param({"guidance_scale": 0}, "guidance_scale", id="guidance_scale_zero", marks=_SKIP_ISSUE_3649),
        pytest.param(
-            {"num_inference_steps": 0}, "num_inference_steps", id="num_inference_steps_zero", marks=_SKIP_ISSUE_3649
+            {"num_inference_steps": 0},
            ("num_inference_steps", "greater than or equal to 1"),
            id="num_inference_steps_zero",
        ),
        pytest.param(
-            {"num_inference_steps": -1}, ("number of steps", "non-negative"), id="num_inference_steps_negative"
+            {"num_inference_steps": -1},
            ("num_inference_steps", "greater than or equal to 1"),
            id="num_inference_steps_negative",
        ),
        pytest.param(
            {"num_inference_steps": 6000},
@@ -157,7 +157,13 @@ def test_speech_missing_required_fields(omni_server: OmniServer, online_client:
                "ref_audio": None,
                "ref_text": None,
            },
-            ("Invalid voice", "vivian"),
+            # NOTE: This is delegating down the TTS adapter, so the message is both specific
            # to the Qwen3TTS adapter and dependent on whether or not the model has any speakers
            # (as opposed to having valida speakers, but getting an invalid speaker). it would
            # be more ideal to validate that this is landing on `.validate()` on the adapter
            # and ensuring that the returned message matches what we would get from calling the
            # adapter validate directly.
            ("no speakers configured",),
            id="customvoice_invalid_voice",
        ),
        pytest.param(
@@ -11,6 +11,7 @@ Tests cover:
 """
 from types import SimpleNamespace
 from unittest.mock import Mock
 import pytest
 import torch
@@ -101,7 +102,12 @@ class TestAttentionSpec:
    def test_block_sparse_defaults_applied_when_backend_selected(self):
        spec = AttentionSpec(backend="RAINFUSION_ATTN")
        assert spec.block_sparse.sparsity == 0.8
-        assert spec.backend_kwargs() == {"sparsity": 0.8, "start_step": 0}
+        assert spec.backend_kwargs() == {
            "sparsity": 0.8,
            "start_step": 0,
            "end_step": 0,
            "precision": "bf16",
        }
    def test_block_sparse_skip_layers_selector_expanded(self):
        spec = AttentionSpec(
@@ -112,6 +118,8 @@ class TestAttentionSpec:
        assert spec.backend_kwargs() == {
            "sparsity": 0.9,
            "start_step": 12,
            "end_step": 0,
            "precision": "bf16",
            "skip_layers": [0, 1, 2, 38],
        }
@@ -121,12 +129,44 @@ class TestAttentionSpec:
    @pytest.mark.parametrize(
        "block_sparse",
-        [{"sparsity": 1.5}, {"start_step": -1}],
+        [{"sparsity": 1.5}, {"start_step": -1}, {"end_step": -1}, {"precision": "bogus"}],
    )
    def test_block_sparse_invalid_values(self, block_sparse):
        with pytest.raises(ValueError):
            AttentionSpec(backend="RAINFUSION_ATTN", block_sparse=block_sparse)
    def test_block_sparse_end_step_precision_forwarded(self):
        spec = AttentionSpec(
            backend="RAINFUSION_ATTN",
            block_sparse={"sparsity": 0.9, "start_step": 5, "end_step": 3, "precision": "fp8"},
        )
        assert spec.backend_kwargs() == {
            "sparsity": 0.9,
            "start_step": 5,
            "end_step": 3,
            "precision": "fp8",
        }
    def test_block_sparse_precision_enum_input_normalized(self):
        """RainFusionPrecision enum members must normalize to their string value.
        Regression: the str,Enum members passed the value check but str() later
        produced "RainFusionPrecision.MIX" instead of "mix" (StrEnum fixes it).
        """
        from vllm_omni.diffusion.attention.backends.rainfusion_attn import RainFusionConfig
        from vllm_omni.diffusion.data import RainFusionPrecision
        spec = AttentionSpec(
            backend="RAINFUSION_ATTN",
            block_sparse={"sparsity": 0.8, "precision": RainFusionPrecision.MIX},
        )
        kwargs = spec.backend_kwargs()
        assert kwargs["precision"] == "mix"
        # The config path is where str(value) used to mangle enum members.
        cfg = RainFusionConfig.from_backend_kwargs(kwargs)
        assert cfg.precision == "mix"
        assert type(cfg.precision) is str
 class TestAttentionConfig:
    def test_empty_config(self):
@@ -427,6 +467,120 @@ class TestCurrentDiffusionConfig:
 class TestAttentionInitUsesCurrentDiffusionConfig:
    @pytest.mark.parametrize("mindiesd_available", [False, True])
    def test_paged_npu_memory_profile_uses_sdpa_only_without_mindiesd(self, monkeypatch, mindiesd_available):
        attention = Attention.__new__(Attention)
        attention._scheduler_paged_kv = True
        attention.paged_kv_cache_role = "primary"
        attention.backend_pref = "FLASH_ATTN"
        attention.attn_backend = SimpleNamespace(supports_piecewise_spans=True, get_name=lambda: "FLASH_ATTN")
        flash_output = torch.ones(1)
        sdpa_output = torch.zeros(1)
        dense_flash = Mock(return_value=flash_output)
        sdpa = Mock(return_value=sdpa_output)
        attention.attention = SimpleNamespace(forward=dense_flash)
        attention.sdpa_fallback = SimpleNamespace(forward=sdpa)
        monkeypatch.setattr(layer_mod, "is_forward_context_available", lambda: True)
        monkeypatch.setattr(
            layer_mod,
            "get_forward_context",
            lambda: SimpleNamespace(in_diffusion_kv_memory_profile=True),
        )
        monkeypatch.setattr(layer_mod.current_omni_platform, "is_npu", lambda: True)
        monkeypatch.setattr(
            layer_mod.current_omni_platform,
            "supports_diffusion_dense_flash_attention",
            lambda: mindiesd_available,
        )
        result = attention._run_local_attention(
            torch.zeros(1, dtype=torch.bfloat16),
            torch.zeros(1, dtype=torch.bfloat16),
            torch.zeros(1, dtype=torch.bfloat16),
            AttentionMetadata(),
        )
        if mindiesd_available:
            assert result is flash_output
            dense_flash.assert_called_once()
            sdpa.assert_not_called()
        else:
            assert result is sdpa_output
            dense_flash.assert_not_called()
            sdpa.assert_called_once()
    def test_dense_flash_import_error_is_not_reclassified(self, monkeypatch):
        attention = Attention.__new__(Attention)
        attention._scheduler_paged_kv = False
        attention.paged_kv_cache_role = "primary"
        attention.backend_pref = "FLASH_ATTN"
        attention.attn_backend = SimpleNamespace(supports_piecewise_spans=True, get_name=lambda: "FLASH_ATTN")
        attention.attention = SimpleNamespace(
            forward=Mock(side_effect=ModuleNotFoundError("No module named 'mindiesd'", name="mindiesd"))
        )
        attention.sdpa_fallback = SimpleNamespace(forward=Mock(side_effect=AssertionError("must not fall back")))
        monkeypatch.setattr(layer_mod, "is_forward_context_available", lambda: False)
        with pytest.raises(ModuleNotFoundError, match="mindiesd"):
            attention._run_local_attention(
                torch.zeros(1, dtype=torch.bfloat16),
                torch.zeros(1, dtype=torch.bfloat16),
                torch.zeros(1, dtype=torch.bfloat16),
                AttentionMetadata(),
            )
    def test_paged_marker_does_not_change_dense_backend(self, monkeypatch):
        class _FakeAttentionImpl:
            def __init__(self, **kwargs):
                self.kwargs = kwargs
            def forward(self, query, key, value, attn_metadata=None):
                return query
        class _DenseBackend:
            supports_paged_kv = False
            @staticmethod
            def get_name() -> str:
                return "TORCH_SDPA"
            @staticmethod
            def get_impl_cls():
                return _FakeAttentionImpl
        calls = []
        def _fake_get_attn_backend_for_role(*args, **kwargs):
            calls.append((args, kwargs))
            return _DenseBackend, None
        monkeypatch.setattr(layer_mod, "get_attn_backend_for_role", _fake_get_attn_backend_for_role)
        monkeypatch.setattr(layer_mod.SDPABackend, "get_impl_cls", staticmethod(lambda: _FakeAttentionImpl))
        monkeypatch.setattr(layer_mod, "build_parallel_attention_strategy", lambda **kwargs: object())
        monkeypatch.setattr(layer_mod, "is_forward_context_available", lambda: False)
        od_config = SimpleNamespace(
            diffusion_attention_config=AttentionConfig(),
            diffusion_kv_mode=layer_mod.DiffusionKVCacheMode.DENSE_LEGACY,
            parallel_config=SimpleNamespace(ring_degree=1),
            diffusion_kv_cache_dtype=None,
            diffusion_kv_cache_skip_step_indices=None,
            diffusion_kv_cache_skip_layer_indices=None,
        )
        with set_current_diffusion_config(od_config):
            attention = Attention(
                num_heads=4,
                head_size=64,
                causal=False,
                softmax_scale=1.0,
                paged_kv_cache_role="primary",
            )
        assert len(calls) == 1
        assert attention.attn_backend is _DenseBackend
        assert attention.backend_pref is None
    def test_attention_init_uses_current_diffusion_config_without_forward_context(self, monkeypatch):
        class _FakeAttentionImpl:
            def __init__(self, **kwargs):
@@ -353,8 +353,15 @@ def test_cross_attn_key_padding_vs_sdpa(k_len):
    print(f"Max absolute difference: {max_diff:.6f}")
    print(f"Mean absolute difference: {mean_diff:.6f}")
-    assert max_diff < 0.01, f"Max difference {max_diff} exceeds threshold 0.01"
+    # Match the platform-specific comparison used by test_fa_vs_sdpa above.
-    assert mean_diff < 0.001, f"Mean difference {mean_diff} exceeds threshold 0.001"
+    if current_omni_platform.is_rocm():
        # AITER FlashAttention and PyTorch SDPA use different BF16 reduction
        # implementations on ROCm, so individual values can differ by one BF16
        # step even though the outputs agree within the expected relative error.
        torch.testing.assert_close(output_fa, output_sdpa, rtol=1e-2, atol=1.5e-2)
    else:
        assert max_diff < 0.01, f"Max difference {max_diff} exceeds threshold 0.01"
        assert mean_diff < 0.001, f"Mean difference {mean_diff} exceeds threshold 0.001"
    print("✓ Case 3 PASSED: FA and SDPA cross-attention outputs are very close!")
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """End-to-end test for ``piecewise_attn`` (CPU).
 Verify that running attention in segments (causal outside full-attn spans,
@@ -98,9 +98,12 @@ def test_paged_piecewise_runner_uses_packed_indices_and_kv_endpoints():
    query = torch.arange(12 * 2 * 4, dtype=torch.float32).reshape(12, 2, 4)
    key = query + 100
    value = query + 200
-    calls = []
+    calls: list[object] = []
-    def run_segment(segment_query, segment_key, segment_value, segment_batch):
+    output_buffer = torch.empty_like(query)
    def run_segment(segment_query, segment_key, segment_value, segment_batch, segment_output_buffer):
        assert segment_output_buffer is None
        calls.append(
            (
                [segment.mode for segment in segment_batch.row_segments],
@@ -119,9 +122,13 @@ def test_paged_piecewise_runner_uses_packed_indices_and_kv_endpoints():
        plan,
        plan.segments,
        run_segment,
        output_buffer=output_buffer,
    )
    assert output is output_buffer
    torch.testing.assert_close(output, query)
    assert all(segment.query_range is None for segment in plan.segments)
    assert not plan.segments_cover_query_contiguously
    assert calls == [
        (["causal", "causal"], [2, 4], [0, 1, 6, 7]),
        (["full", "full"], [5, 7], [2, 3, 4, 8, 9, 10]),
@@ -129,6 +136,161 @@ def test_paged_piecewise_runner_uses_packed_indices_and_kv_endpoints():
    ]
 def test_paged_piecewise_runner_preserves_homogeneous_batch_without_scatter(monkeypatch: pytest.MonkeyPatch):
    plan = build_paged_piecewise_plan(
        [[(2, 5)], [(2, 5)]],
        query_offsets=[0, 0],
        query_lens=[6, 6],
        seq_lens=[6, 6],
        device=DEVICE,
    )
    assert plan.homogeneous_batch_shape == (2, 6)
    query = torch.arange(12 * 2 * 4, dtype=torch.float32).reshape(12, 2, 4)
    calls: list[torch.Tensor] = []
    def fail_index_copy(*_args, **_kwargs):
        raise AssertionError("homogeneous piecewise path must not use index_copy_")
    monkeypatch.setattr(torch.Tensor, "index_copy_", fail_index_copy)
    def run_segment(segment_query, segment_key, segment_value, _metadata, segment_output_buffer):
        assert segment_key is None
        assert segment_value is None
        assert segment_output_buffer is None
        calls.append(segment_query.clone())
        return segment_query + 1
    output_buffer = torch.empty_like(query)
    output = run_paged_piecewise_plan(
        query,
        None,
        None,
        plan,
        plan.segments,
        run_segment,
        output_buffer=output_buffer,
        use_homogeneous_batch=True,
    )
    assert output is output_buffer
    torch.testing.assert_close(output, query + 1)
    assert [call.shape[0] for call in calls] == [4, 6, 2]
    torch.testing.assert_close(calls[0], query[[0, 1, 6, 7]])
    torch.testing.assert_close(calls[1], query[[2, 3, 4, 8, 9, 10]])
    torch.testing.assert_close(calls[2], query[[5, 11]])
 def test_paged_piecewise_runner_keeps_indexed_fallback_for_heterogeneous_layout():
    plan = build_paged_piecewise_plan(
        [[(2, 5)], [(1, 4)]],
        query_offsets=[0, 0],
        query_lens=[6, 6],
        seq_lens=[6, 6],
        device=DEVICE,
    )
    assert plan.homogeneous_batch_shape is None
    query = torch.arange(12 * 2 * 4, dtype=torch.float32).reshape(12, 2, 4)
    output = run_paged_piecewise_plan(
        query,
        None,
        None,
        plan,
        plan.segments,
        lambda segment_query, *_args: segment_query,
        use_homogeneous_batch=True,
    )
    torch.testing.assert_close(output, query)
 def test_paged_piecewise_runner_slices_contiguous_segments():
    plan = build_paged_piecewise_plan(
        [[(2, 5)]],
        query_offsets=[0],
        query_lens=[6],
        seq_lens=[6],
        device=DEVICE,
    )
    query = torch.arange(6 * 2 * 4, dtype=torch.float32).reshape(6, 2, 4)
    key = query + 100
    value = query + 200
    expected_ranges = [(0, 2), (2, 5), (5, 6)]
    calls: list[tuple[int, int]] = []
    def run_segment(segment_query, segment_key, segment_value, _segment_batch, segment_output_buffer):
        assert segment_output_buffer is None
        expected_range = expected_ranges[len(calls)]
        start, end = expected_range
        calls.append(expected_range)
        assert segment_query.untyped_storage().data_ptr() == query.untyped_storage().data_ptr()
        assert segment_key.untyped_storage().data_ptr() == key.untyped_storage().data_ptr()
        assert segment_value.untyped_storage().data_ptr() == value.untyped_storage().data_ptr()
        torch.testing.assert_close(segment_query, query[start:end])
        return segment_query
    output = run_paged_piecewise_plan(query, key, value, plan, plan.segments, run_segment)
    assert calls == expected_ranges
    assert [segment.query_range for segment in plan.segments] == expected_ranges
    assert plan.segments_cover_query_contiguously
    torch.testing.assert_close(output, query)
 def test_paged_piecewise_runner_writes_contiguous_segments_into_output_buffer():
    plan = build_paged_piecewise_plan(
        [[(2, 5)]],
        query_offsets=[0],
        query_lens=[6],
        seq_lens=[6],
        device=DEVICE,
    )
    query = torch.arange(6 * 2 * 4, dtype=torch.float32).reshape(6, 2, 4)
    output_buffer = torch.empty(6, 2, 3)
    targets = []
    def run_segment(segment_query, _key, _value, _metadata, segment_output_buffer):
        assert segment_output_buffer is not None
        assert segment_output_buffer.untyped_storage().data_ptr() == output_buffer.untyped_storage().data_ptr()
        targets.append(segment_output_buffer)
        return segment_output_buffer.copy_(segment_query[..., :3])
    output = run_paged_piecewise_plan(
        query,
        query,
        query,
        plan,
        plan.segments,
        run_segment,
        output_buffer=output_buffer,
    )
    assert output is output_buffer
    assert [target.shape[0] for target in targets] == [2, 3, 1]
    torch.testing.assert_close(output, query[..., :3])
 def test_paged_piecewise_runner_returns_single_full_segment_directly():
    plan = build_paged_piecewise_plan(
        [[]],
        query_offsets=[0],
        query_lens=[6],
        seq_lens=[6],
        device=DEVICE,
    )
    query = torch.randn(6, 2, 4)
    runner_output = query + 1
    output = run_paged_piecewise_plan(
        query,
        query,
        query,
        plan,
        plan.segments,
        lambda *_args: runner_output,
    )
    assert output is runner_output
 def test_paged_piecewise_runner_uses_runner_output_shape():
    plan = build_paged_piecewise_plan(
        [[(2, 5)]],
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Geometry contract RAINFUSION_ATTN enforces before handing a forward to rf_v2.
@@ -11,18 +11,25 @@ to the always-kept prefix segment before generating the block mask.
 """
 import dataclasses
 import sys
 import types
 from unittest import mock
 import pytest
 import torch
-from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata, VideoTokenLayout
+from vllm_omni.diffusion.attention.backends import rainfusion_attn
 from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata, VideoTokenLayout, VideoTokenSpan
 from vllm_omni.diffusion.attention.backends.rainfusion_attn import (
    _BLOCK_SIZE,
    RainFusionAttentionBackend,
    RainFusionAttentionImpl,
    RainFusionPlan,
 )
 from vllm_omni.platforms import current_omni_platform
 pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.npu]
 PREFIX_ROWS = 710  # 14 text rows + 696 audio rows
 ALIGNED_GRID = (62, 24, 40)  # 1280x768 -> 59520 video rows, 465 blocks
 MISALIGNED_GRID = (62, 24, 42)  # 1344x768 -> 62496 video rows, 488 blocks + 32 rows
@@ -97,6 +104,62 @@ def test_video_segment_must_be_the_tail_of_packed_document_zero():
    assert make_impl()._resolve_plan(metadata) is None
 def test_ref2va_multi_video_spans_resolve():
    layout = VideoTokenLayout(
        used_len=12000,
        video_spans=(
            VideoTokenSpan(start=128, latent_grid=(4, 16, 64), role="reference"),
            VideoTokenSpan(start=5000, latent_grid=(4, 16, 64), role="target"),
        ),
    )
    plan = make_impl()._resolve_plan(AttentionMetadata(extra={"max_seqlen_q": 12000}, video_layout=layout))
    assert plan is not None
    assert plan.used_len == 12000
    assert plan.video_spans == [
        {"start": 128, "latent_shape": [4, 16, 64]},
        {"start": 5000, "latent_shape": [4, 16, 64]},
    ]
 def test_invalid_ref2va_video_spans_fall_back_to_dense():
    layout = VideoTokenLayout(
        used_len=12000,
        video_spans=(
            VideoTokenSpan(start=128, latent_grid=(4, 16, 64), role="reference"),
            VideoTokenSpan(start=4000, latent_grid=(4, 16, 64), role="target"),
        ),
    )
    assert make_impl()._resolve_plan(AttentionMetadata(extra={"max_seqlen_q": 12000}, video_layout=layout)) is None
 def test_validate_available_rejects_legacy_mindiesd(monkeypatch):
    mindiesd = types.ModuleType("mindiesd")
    def sparse_attention(query, key, value, **kwargs):
        return query
    mindiesd.sparse_attention = sparse_attention
    monkeypatch.setitem(sys.modules, "mindiesd", mindiesd)
    monkeypatch.setattr("importlib.util.find_spec", lambda _: object())
    with pytest.raises(ValueError, match="video_spans"):
        RainFusionAttentionBackend.validate_available()
 def test_validate_available_accepts_new_mindiesd(monkeypatch):
    mindiesd = types.ModuleType("mindiesd")
    def sparse_attention(query, key, value, *, video_spans=None, **kwargs):
        return query
    mindiesd.sparse_attention = sparse_attention
    monkeypatch.setitem(sys.modules, "mindiesd", mindiesd)
    monkeypatch.setattr("importlib.util.find_spec", lambda _: object())
    RainFusionAttentionBackend.validate_available()
@pytest.mark.parametrize("grid", [(4, 24, 40), (1, 24, 40)])
 def test_short_video_stays_dense(grid):
    assert make_impl()._resolve_plan(make_metadata(grid)) is None
@@ -169,7 +232,9 @@ def test_irregular_video_tail_reproduces_dense_attention_with_updated_mindiesd()
    heads, head_dim = 4, 128
    q, k, v = (torch.randn(1, used, heads, head_dim, dtype=torch.bfloat16, device="npu") for _ in range(3))
-    impl = make_impl()
+    # Explicit bf16: the dense comparison exercises the plain sparse kernel,
    # and older MindIE-SD releases do not accept precision= (gate would raise).
    impl = make_impl(precision="bf16")
    impl.rainfusion = dataclasses.replace(impl.rainfusion, sparsity=0.0)
    plan = RainFusionPlan(prefix_len=prefix_len, used_len=used, latent_shape=list(grid))
    out = impl._forward_sparse_npu(q, k, v, plan)
@@ -182,3 +247,90 @@ def test_irregular_video_tail_reproduces_dense_attention_with_updated_mindiesd()
    ).transpose(1, 2)
    error = (out.float() - reference.float()).abs().mean() / reference.float().abs().mean()
    assert error < 2e-3, f"mean relative error {error:.4%} against dense attention"
 # end_step tail fallback: the last ``end_step`` denoise steps must stay dense.
 def test_end_step_tail_window_stays_dense():
    """step_idx inside the last end_step window must resolve no sparse plan."""
    impl = make_impl(end_step=3)
    fc = mock.Mock()
    fc.denoise_step_idx = 47
    fc.total_denoise_steps = 50
    with (
        mock.patch.object(rainfusion_attn, "is_forward_context_available", return_value=True),
        mock.patch.object(rainfusion_attn, "get_forward_context", return_value=fc),
    ):
        assert impl._resolve_plan(make_metadata(ALIGNED_GRID)) is None
 def test_end_step_before_tail_window_still_sparse():
    """step_idx before the tail window must still resolve a sparse plan."""
    impl = make_impl(end_step=3)
    fc = mock.Mock()
    fc.denoise_step_idx = 46
    fc.total_denoise_steps = 50
    with (
        mock.patch.object(rainfusion_attn, "is_forward_context_available", return_value=True),
        mock.patch.object(rainfusion_attn, "get_forward_context", return_value=fc),
    ):
        assert impl._resolve_plan(make_metadata(ALIGNED_GRID)) is not None
 def test_end_step_zero_never_triggers_tail_fallback():
    """end_step=0 must not fall back even on the final denoise step."""
    impl = make_impl(end_step=0)
    fc = mock.Mock()
    fc.denoise_step_idx = 49
    fc.total_denoise_steps = 50
    with (
        mock.patch.object(rainfusion_attn, "is_forward_context_available", return_value=True),
        mock.patch.object(rainfusion_attn, "get_forward_context", return_value=fc),
    ):
        assert impl._resolve_plan(make_metadata(ALIGNED_GRID)) is not None
 # precision capability gate: a mindiesd without sparse_attention(precision=)
 # must fail fast instead of silently running the BF16 path.
 def _fake_mindiesd_module():
    """Install a minimal stand-in so the import inside _forward_sparse_npu succeeds."""
    import types
    fake = types.ModuleType("mindiesd")
    fake.sparse_attention = lambda *args, **kwargs: None
    return fake
 def test_precision_non_bf16_requires_mindiesd_support():
    """precision != bf16 against a mindiesd lacking the kwarg must raise RuntimeError."""
    import sys
    impl = make_impl(precision="mix")
    sys.modules["mindiesd"] = _fake_mindiesd_module()
    try:
        with mock.patch.object(rainfusion_attn, "_mindiesd_supports_precision", return_value=False):
            with pytest.raises(RuntimeError, match="requires MindIE-SD"):
                impl._forward_sparse_npu(None, None, None, None)
    finally:
        sys.modules.pop("mindiesd", None)
 def test_precision_non_bf16_passes_gate_when_supported():
    """precision != bf16 with a capable mindiesd must not raise from the gate."""
    import sys
    impl = make_impl(precision="mix")
    sys.modules["mindiesd"] = _fake_mindiesd_module()
    try:
        with mock.patch.object(rainfusion_attn, "_mindiesd_supports_precision", return_value=True):
            # q/k/v shapes: [B, S, N, D]; plan geometry must match S.
            q = torch.randn(1, 8, 4, 128)
            plan = RainFusionPlan(prefix_len=0, used_len=8, latent_shape=[2, 2, 2])
            # The gate must pass; the fake mindiesd returns None so no crash.
            out = impl._forward_sparse_npu(q, q, q, plan)
            assert out is None
    finally:
        sys.modules.pop("mindiesd", None)
@@ -7,7 +7,7 @@ End-to-end test for diffusion batching via AsyncOmni.
 This test fires multiple concurrent ``AsyncOmni.generate()`` calls for a
 diffusion model and validates that every caller receives its correct
 individual result.  When the underlying diffusion stage is configured with
-``batch_size > 1`` (via stage config or ``StageDiffusionClient``), the
+``max_num_seqs > 1``, the
 requests will be batched internally.
 Even without explicit batching config this test is useful for verifying
@@ -260,7 +260,7 @@ async def compare_single_vs_parallel(
 ) -> None:
    """Run the same prompts sequentially then in parallel and print a comparison."""
-    omni = AsyncOmni(model=model, max_num_seqs=batch_size, diffusion_batch_size=batch_size)
+    omni = AsyncOmni(model=model, max_num_seqs=batch_size)
    try:
        await warmup(omni, WARMUP_PROMPTS)
        single_time = await run_single(omni, prompts)
@@ -288,7 +288,7 @@ async def main(model: str, num_prompts: int, mode: str, batch_size: int = 1) ->
        await compare_single_vs_parallel(model, prompts, batch_size=batch_size)
        return
-    omni = AsyncOmni(model=model, max_num_seqs=batch_size, diffusion_batch_size=batch_size)
+    omni = AsyncOmni(model=model, max_num_seqs=batch_size)
    try:
        await warmup(omni, WARMUP_PROMPTS)
@@ -389,7 +389,7 @@ def test_diffusion_batching_async_concurrent(model_name: str):
    fired via asyncio.gather. Each request_id must appear in the results."""
    async def _inner():
-        omni = AsyncOmni(model=model_name, diffusion_batch_size=1)
+        omni = AsyncOmni(model=model_name, max_num_seqs=1)
        try:
            prompts = TEST_PROMPTS[:4]
            sp = _default_sampling_params()
@@ -430,7 +430,7 @@ def test_diffusion_batching_list_prompt_rejected(model_name: str):
    """
    async def _inner():
-        omni = AsyncOmni(model=model_name, diffusion_batch_size=4)
+        omni = AsyncOmni(model=model_name, max_num_seqs=4)
        try:
            prompts = TEST_PROMPTS[:4]
            sp = _default_sampling_params()
@@ -528,7 +528,7 @@ if __name__ == "__main__":
        "--batch-size",
        type=int,
        default=1,
-        help="Scheduler in-flight width (--max-num-seqs) and client diffusion_batch_size",
+        help="Scheduler in-flight width (--max-num-seqs)",
    )
    parser.add_argument(
        "--mode",
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """
 Unit tests for TeaCache extractor functions.
@@ -18,6 +18,7 @@ Currently implemented:
 - TestMiniMaxH3Extractor: MiniMaxH3DiTModel extractor
 """
 import math
 from abc import ABC, abstractmethod
 from types import SimpleNamespace
 from unittest.mock import MagicMock, Mock, patch
@@ -650,6 +651,29 @@ class TestMiniMaxH3Extractor(BaseExtractorTest):
        torch.testing.assert_close(actual_video, expected_video)
        torch.testing.assert_close(actual_audio, expected_audio)
    def test_extractor_uses_prepared_rope_table(
        self,
        minimax_h3_module,
        sample_inputs,
        monkeypatch,
    ):
        """TeaCache must consume the denoise-loop RoPE table without rebuilding it."""
        rope_table = minimax_h3_module.prepare_rope_table(
            sample_inputs["img_position_ids"],
            seq_len=sample_inputs["x"].shape[1],
        )
        def unexpected_rope_build(*args, **kwargs):
            raise AssertionError("TeaCache extractor rebuilt the prepared RoPE table")
        monkeypatch.setattr(minimax_h3_module, "prepare_rope_table", unexpected_rope_build)
        monkeypatch.setattr(minimax_h3_module.rope, "forward", unexpected_rope_build)
        context = extract_minimax_h3_context(
            minimax_h3_module,
            **{**sample_inputs, "rope_table": rope_table},
        )
        context.run_transformer_blocks()
    def test_run_transformer_blocks_forwards_packed_layout(
        self,
        minimax_h3_module,
@@ -705,6 +729,222 @@ class TestMiniMaxH3Extractor(BaseExtractorTest):
        assert video_logits.shape[0] == sample_inputs["img_pos_info"]["position_ids"].shape[0]
        assert audio_logits.shape[0] == sample_inputs["audio_pos_info"]["position_ids"].shape[0]
    @pytest.mark.sp
    def test_teacache_strict_sp_keeps_cache_state_rank_local(
        self,
        minimax_h3_module,
        sample_inputs,
        monkeypatch,
    ):
        """Strict SP must not mix local TeaCache state with gathered block rows."""
        from vllm_omni.diffusion.attention.ops import minimax_h3_modulation
        from vllm_omni.diffusion.cache.teacache.config import TeaCacheConfig
        from vllm_omni.diffusion.cache.teacache.hook import TeaCacheHook
        from vllm_omni.diffusion.distributed import parallel_state
        seq_len = sample_inputs["x"].shape[1]
        local_len = seq_len // 2
        hidden_size = minimax_h3_module.hidden_size
        video_width = minimax_h3_module.arch.latents_dim * math.prod(minimax_h3_module.arch.patch_size)
        audio_width = minimax_h3_module.arch.audio_latents_dim
        captured_modulation_indices: list[torch.Tensor] = []
        class _StrictSPPrepare(nn.Module):
            def __init__(self):
                super().__init__()
                self.global_indices: list[torch.Tensor] = []
            def forward(self, hidden, rope_table, combined_indices):
                assert hidden.shape[0] == local_len
                assert rope_table.shape[0] == local_len
                self.global_indices.append(combined_indices.clone())
                return hidden, rope_table, combined_indices.narrow(0, 0, local_len)
        class _StrictSPGather(nn.Module):
            def __init__(self):
                super().__init__()
                self.inputs: list[torch.Tensor] = []
                self.outputs: list[torch.Tensor] = []
            def forward(self, tensor):
                self.inputs.append(tensor.clone())
                if tensor.shape[-1] == hidden_size:
                    raise AssertionError("TeaCache must not gather hidden states in strict SP")
                assert tensor.shape == (local_len, video_width + audio_width)
                # Model the all-gather of compact logits from every SP rank.
                # postprocess() indexes the resulting global layout with
                # infer_out_pos/audio_pos, including rows owned by another rank.
                gathered = tensor.repeat(seq_len // local_len, 1)
                assert gathered.shape == (seq_len, video_width + audio_width)
                self.outputs.append(gathered)
                return gathered
        class _StrictSPGroup:
            def __init__(self):
                super().__init__()
                self.local_decisions: list[int] = []
                self.ops = []
            def all_reduce(self, decision, op):
                self.local_decisions.append(int(decision.item()))
                self.ops.append(op)
                # Simulate another rank requiring the second step to execute
                # the collective block path even though this rank would cache.
                if len(self.local_decisions) == 2:
                    decision.fill_(1)
                return decision
        class _StrictSPBlock(nn.Module):
            def __init__(self):
                super().__init__()
                self.block_indices: list[torch.Tensor] = []
            def adaln_proj(self, t_emb):
                del t_emb
                zeros = torch.zeros(3, hidden_size, dtype=torch.float32)
                return zeros, zeros
            def norm1(self, hidden):
                return hidden.clone()
            def forward(self, hidden, *, combined_indices, rope_table, **kwargs):
                del kwargs
                assert hidden.shape[0] == local_len
                assert rope_table.shape[0] == local_len
                self.block_indices.append(combined_indices.clone())
                return hidden + 1
        class _StrictSPFinalLayer(nn.Module):
            def forward(self, hidden, *, t_emb, inverse_indices):
                del t_emb, inverse_indices
                assert hidden.shape == (local_len, hidden_size)
                return hidden[:, :video_width].float(), hidden[:, :audio_width].float()
        local_prepare = _StrictSPPrepare()
        gather = _StrictSPGather()
        sp_group = _StrictSPGroup()
        block = _StrictSPBlock()
        minimax_h3_module.local_sp_prepare = local_prepare
        minimax_h3_module.sp_gather = gather
        minimax_h3_module.blocks = nn.ModuleList([block])
        minimax_h3_module.final_layer = _StrictSPFinalLayer()
        monkeypatch.setattr(minimax_h3_module, "_rope_local_span", lambda _seq_len: (0, local_len))
        monkeypatch.setattr(
            minimax_h3_module,
            "_embed",
            lambda **kwargs: (
                torch.arange(local_len * hidden_size, dtype=torch.float32).reshape(local_len, hidden_size),
                torch.zeros(3, hidden_size, dtype=torch.float32),
            ),
        )
        monkeypatch.setattr(
            minimax_h3_module,
            "prepare_rope_table",
            lambda *args, **kwargs: torch.zeros(local_len, 6, dtype=torch.float32),
        )
        monkeypatch.setattr(parallel_state, "get_sp_group", lambda: sp_group)
        original_indexed_scale_shift = minimax_h3_modulation.indexed_scale_shift_
        def capture_indexed_scale_shift(hidden, shift, scale, indices):
            captured_modulation_indices.append(indices.clone())
            return original_indexed_scale_shift(hidden, shift, scale, indices)
        monkeypatch.setattr(minimax_h3_modulation, "indexed_scale_shift_", capture_indexed_scale_shift)
        local_img_pos = torch.tensor([0, 1])
        local_audio_pos = torch.tensor([2, 3])
        token_tags = torch.full((seq_len,), -1, dtype=torch.long)
        token_tags[local_img_pos] = 0
        token_tags[local_audio_pos] = 2
        inverse_indices = torch.zeros(seq_len, dtype=torch.long)
        inputs = {
            **sample_inputs,
            "inverse_indices": inverse_indices,
            "token_tags": token_tags,
            "update_mask": torch.ones(local_img_pos.numel(), dtype=torch.bool),
            "prompt_embeds": torch.empty(0, 6),
            "img_pos_info": {"position_ids": local_img_pos},
            "audio_pos_info": {"position_ids": local_audio_pos},
            "text_pos_info": {"position_ids": torch.empty(0, dtype=torch.long)},
            "img_pos_for_infer_output_info": {"position_ids": local_img_pos},
            "refiner_packed_seq_params": {
                "cu_seqlens_q": torch.tensor([0, 0], dtype=torch.int32),
                "max_seqlen_q": 0,
            },
        }
        expected_local_indices = (inverse_indices * 3 + token_tags.clamp(min=0))[:local_len]
        hook = TeaCacheHook(
            TeaCacheConfig(
                transformer_type="MiniMaxH3DiTModel",
                rel_l1_thresh=0.3,
            )
        )
        hook.initialize_hook(minimax_h3_module)
        hook.new_forward(minimax_h3_module, **inputs)
        hook.new_forward(minimax_h3_module, **inputs)
        video_logits, audio_logits = hook.new_forward(minimax_h3_module, **inputs)
        state = hook.state_manager.get_state()
        assert state.previous_residual is not None
        assert state.previous_residual.shape == (local_len, hidden_size)
        assert len(captured_modulation_indices) == 3
        assert all(torch.equal(indices, expected_local_indices) for indices in captured_modulation_indices)
        assert sp_group.local_decisions == [1, 0, 0]
        assert len(sp_group.ops) == 3
        assert all(op == torch.distributed.ReduceOp.MAX for op in sp_group.ops)
        assert len(local_prepare.global_indices) == 2
        assert all(indices.shape == (seq_len,) for indices in local_prepare.global_indices)
        assert len(block.block_indices) == 2
        assert all(torch.equal(indices, expected_local_indices) for indices in block.block_indices)
        assert len(gather.inputs) == 3
        assert video_logits.shape == (local_img_pos.numel(), video_width)
        assert audio_logits.shape == (local_audio_pos.numel(), audio_width)
    @pytest.mark.cpu
    def test_teacache_non_strict_sp_gathers_hidden_once(
        self,
        minimax_h3_module,
        sample_inputs,
        monkeypatch,
    ):
        """The non-strict SP cache path must not gather global hidden rows twice."""
        from vllm_omni.diffusion.cache.teacache.config import TeaCacheConfig
        from vllm_omni.diffusion.cache.teacache.hook import TeaCacheHook
        class _Prepare(nn.Module):
            def forward(self, hidden, rope_table, combined_indices):
                return hidden, rope_table, combined_indices
        class _Gather(nn.Module):
            def __init__(self):
                super().__init__()
                self.inputs: list[torch.Tensor] = []
            def forward(self, tensor):
                self.inputs.append(tensor.clone())
                return tensor
        seq_len = sample_inputs["x"].shape[1]
        gather = _Gather()
        minimax_h3_module.sp_prepare = _Prepare()
        minimax_h3_module.sp_gather = gather
        monkeypatch.setattr(minimax_h3_module, "_rope_local_span", lambda _seq_len: (0, seq_len))
        hook = TeaCacheHook(
            TeaCacheConfig(
                transformer_type="MiniMaxH3DiTModel",
                rel_l1_thresh=0.3,
            )
        )
        hook.initialize_hook(minimax_h3_module)
        hook.new_forward(minimax_h3_module, **sample_inputs)
        assert len(gather.inputs) == 1
        assert gather.inputs[0].shape == (seq_len, minimax_h3_module.hidden_size)
    def test_invalid_module_raises_error(self):
        """Test that invalid module without blocks raises ValueError."""
        invalid_module = Mock()
@@ -0,0 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import pytest
 import torch
@pytest.fixture(autouse=True)
 def restore_torch_default_dtype():
    """Keep process-global dtype changes from leaking between diffusion tests."""
    default_dtype = torch.get_default_dtype()
    try:
        yield
    finally:
        torch.set_default_dtype(default_dtype)
@@ -9,6 +9,7 @@ import torch
 from diffusers import DiffusionPipeline  # pyright: ignore[reportPrivateImportUsage]
 from PIL import Image
 import vllm_omni.diffusion.data as diffusion_data
 from tests.helpers.mark import hardware_test
 from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient, dummy_messages_from_mix_data
 from vllm_omni.diffusion.data import (
@@ -629,6 +630,46 @@ class TestPipelineArgumentsHandling:
        with pytest.raises(ValueError):
            pipeline.forward(DiffusionRequestBatch(requests=[problematic_request]))
        interpolation_request = _make_request(
            sampling_params=OmniDiffusionSamplingParams(
                enable_frame_interpolation=True,
            ),
        )
        with pytest.raises(ValueError, match="Frame interpolation is not supported with the Diffusers backend"):
            pipeline.forward(DiffusionRequestBatch(requests=[interpolation_request]))
    def test_diffusers_config_preserves_checkpoint_class_for_metadata(self, mocker):
        class MockWanImageToVideoPipeline:
            pass
        od_config = _make_od_config(model_class_name=None)
        mocker.patch(
            "vllm_omni.diffusion.utils.hf_utils.get_diffusion_model_index",
            return_value={"_class_name": "WanImageToVideoPipeline"},
        )
        mocker.patch.object(
            diffusion_data,
            "diffusers",
            SimpleNamespace(WanImageToVideoPipeline=MockWanImageToVideoPipeline),
        )
        mocker.patch.object(OmniDiffusionConfig, "update_multimodal_support")
        od_config.enrich_config()
        assert od_config.model_class_name == "WanImageToVideoPipeline"
        assert od_config.diffusers_pipeline_cls is MockWanImageToVideoPipeline
    def test_diffusers_config_falls_back_to_adapter_without_pipeline_index(self, mocker):
        od_config = _make_od_config(model_class_name=None)
        mocker.patch(
            "vllm_omni.diffusion.utils.hf_utils.get_diffusion_model_index",
            return_value=None,
        )
        od_config.enrich_config()
        assert od_config.model_class_name == "DiffusersAdapterPipeline"
@pytest.mark.advanced_model
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 from __future__ import annotations
@@ -485,8 +485,16 @@ def test_valid_sequence_and_context_install_into_native_rows(monkeypatch: pytest
    assert runner.get_diffusion_kv_row("req-0", 0, "text") == 1
    sequence_binding = runner._resolve_paged_attention_row("req-0", 0, None)
    context_binding = runner._resolve_paged_attention_row("req-0", None, "text")
-    assert (sequence_binding.row_index, sequence_binding.max_seq_len) == (0, 9)
+    assert (sequence_binding.row_index, sequence_binding.max_seq_len, sequence_binding.block_ids) == (
-    assert (context_binding.row_index, context_binding.max_seq_len) == (1, 5)
+        0,
        9,
        ((1, 2, 3), (4, 5)),
    )
    assert (context_binding.row_index, context_binding.max_seq_len, context_binding.block_ids) == (
        1,
        5,
        ((6, 7), (8,)),
    )
 def test_block_table_mutations_invalidate_prepared_attention_batches(
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 import pytest
@@ -36,7 +36,7 @@ def test_non_mapping_omni_kv_config_is_rejected() -> None:
        OmniDiffusionConfig.from_kwargs(omni_kv_config="paged_scheduler")
-def test_paged_scheduler_rejects_dense_legacy_kv_receive() -> None:
+def test_paged_scheduler_rejects_dense_connector_kv_receive() -> None:
    with pytest.raises(ValueError, match="does not support imported AR KV"):
        OmniDiffusionConfig.from_kwargs(
            diffusion_kv_mode="paged_scheduler",
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 from types import SimpleNamespace
 from unittest.mock import Mock
@@ -290,6 +290,7 @@ def test_minus_one_max_model_len_is_auto_fitted_by_native_cache_sizing() -> None
        dtype=torch.bfloat16,
        non_causal=True,
    )
    model_limit = vllm_config.model_config.max_model_len
    initialization.build_native_kv_cache_configs(
        vllm_config,
@@ -298,4 +299,8 @@ def test_minus_one_max_model_len_is_auto_fitted_by_native_cache_sizing() -> None
    )
    assert vllm_config.model_config.original_max_model_len == -1
-    assert vllm_config.model_config.max_model_len == 256
+    # Sixteen physical pages provide ``16 * block_size`` tokens.  The native
    # backend chooses the block geometry, so keep this assertion backend
    # independent instead of assuming the generic 16-token block size.
    expected_max_model_len = min(model_limit, spec.block_size * 16)
    assert vllm_config.model_config.max_model_len == expected_max_model_len
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 from __future__ import annotations
@@ -20,12 +20,16 @@ from vllm_omni.diffusion.attention.layer import Attention
 from vllm_omni.diffusion.diffusion_kv import paged_attention_adapter as adapter_module
 from vllm_omni.diffusion.diffusion_kv.paged_attention_adapter import (
    DiffusionPagedAttentionAdapter,
    DiffusionPagedAttentionMetadata,
    DiffusionPagedAttentionRow,
    DiffusionPagedAttentionRowBinding,
    DiffusionPagedAttentionRuntime,
 )
 from vllm_omni.diffusion.forward_context import (
    get_forward_context,
    override_paged_kv_adapter,
    set_forward_context,
    set_forward_context_denoise_step_idx,
 )
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
@@ -122,8 +126,8 @@ class _FakeNativeImpl:
    def do_kv_cache_update(
        self,
        _layer,
-        key: torch.Tensor,
+        key: torch.Tensor | None,
-        value: torch.Tensor,
+        value: torch.Tensor | None,
        _kv_cache,
        slot_mapping: torch.Tensor,
    ) -> None:
@@ -155,6 +159,11 @@ def _make_adapter(
    capacity: int = 16,
 ) -> tuple[DiffusionPagedAttentionAdapter, _FakeBlockTables, _FakeLayer, list[tuple]]:
    events: list[tuple] = []
    monkeypatch.setattr(
        adapter_module.current_omni_platform,
        "requires_diffusion_paged_kv_prewrite",
        lambda: False,
    )
    block_tables = _FakeBlockTables()
    block_tables.max_num_batched_tokens = capacity
    block_tables.num_blocks.np.fill(math.ceil(capacity / block_tables.block_sizes[0]))
@@ -301,6 +310,47 @@ def test_paged_adapter_accepts_native_gqa_shapes(monkeypatch: pytest.MonkeyPatch
    assert context.value_write.shape == (3, 8, 4)
 def test_paged_adapter_preserves_strided_kv_for_platform_writer(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch)
    batch = adapter.prepare_batch(
        [DiffusionPagedAttentionRow(request_id="req-0", sequence_id=0, query_len=5, seq_len=5)]
    )
    packed_qkv = torch.randn(1, 5, 3 * 2 * 4)
    query, key, value = packed_qkv.split(2 * 4, dim=-1)
    query = query.reshape(1, 5, 2, 4)
    key = key.reshape(1, 5, 2, 4)
    value = value.reshape(1, 5, 2, 4)
    assert not key.is_contiguous()
    assert not value.is_contiguous()
    with adapter.activate(batch):
        context = adapter.prepare_layer_context("layer-0", query, key, value)
    assert context.query.is_contiguous()
    assert not context.key_write.is_contiguous()
    assert not context.value_write.is_contiguous()
    assert context.key_write.data_ptr() == key.data_ptr()
    assert context.value_write.data_ptr() == value.data_ptr()
 def test_paged_adapter_packs_noncontiguous_kv_head_layout(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch)
    batch = adapter.prepare_batch(
        [DiffusionPagedAttentionRow(request_id="req-0", sequence_id=0, query_len=5, seq_len=5)]
    )
    query = torch.randn(1, 5, 2, 4)
    key = torch.randn(1, 5, 4, 2).transpose(-1, -2)
    value = torch.randn(1, 5, 4, 2).transpose(-1, -2)
    assert key.shape == (1, 5, 2, 4)
    assert key.stride(-1) != 1
    with adapter.activate(batch):
        context = adapter.prepare_layer_context("layer-0", query, key, value)
    assert context.key_write.is_contiguous()
    assert context.value_write.is_contiguous()
 def test_omni_paged_backend_defers_backend_owned_cache_update(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, layer, _ = _make_adapter(monkeypatch)
    layer.attn_backend.forward_includes_kv_cache_update = True
@@ -350,10 +400,48 @@ def test_omni_piecewise_backend_defers_backend_owned_cache_update(monkeypatch: p
    assert [metadata.query_start_loc_cpu.tolist() for metadata in segment_metadata] == [[0, 1], [0, 2], [0, 1]]
    assert [metadata.positions.tolist() for metadata in segment_metadata] == [[0], [1, 2], [3]]
    assert [metadata.slot_mappings.tolist() for metadata in segment_metadata] == [[[0]], [[1, 2]], [[3]]]
    for event in events[1:]:
        build_kwargs = event[1]
        assert build_kwargs["positions"].untyped_storage().data_ptr() == batch.positions.untyped_storage().data_ptr()
        assert (
            build_kwargs["slot_mappings"].untyped_storage().data_ptr()
            == batch.slot_mappings.untyped_storage().data_ptr()
        )
 def test_platform_prewrite_updates_once_then_piecewise_reads_cache(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, layer, _ = _make_adapter(monkeypatch)
    layer.attn_backend.forward_includes_kv_cache_update = True
    monkeypatch.setattr(
        adapter_module.current_omni_platform,
        "requires_diffusion_paged_kv_prewrite",
        lambda: True,
    )
    adapter.resolve_row = lambda *_args: DiffusionPagedAttentionRowBinding(
        row_index=2,
        max_seq_len=16,
        block_ids=((3, 7, 9, 11),),
    )
    row = DiffusionPagedAttentionRow(request_id="req-0", sequence_id=0, query_len=4, seq_len=4)
    batch = adapter.prepare_batch([row])
    qkv = torch.randn(1, 4, 2, 4)
    metadata = SimpleNamespace(full_attn_spans=[[(1, 3)]], extra={})
    with adapter.activate(batch):
        context = adapter.prepare_layer_context("layer-0", qkv, qkv, qkv, omni_attn_metadata=metadata)
        output = _run_omni_paged_backend(context)
    assert torch.equal(output, qkv)
    assert layer.native_events == ["update", "forward", "forward", "forward"]
    assert len(layer.updates) == 1
    assert all(call[1] is None and call[2] is None for call in layer.calls)
 def test_omni_paged_backend_runs_hunyuan_piecewise_segments(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, layer, events = _make_adapter(monkeypatch)
    # Exercise the Ascend-only homogeneous batch path. CUDA keeps its native
    # output-buffer contract and is covered by the GPU adapter tests.
    monkeypatch.setattr(adapter_module.current_omni_platform, "is_npu", lambda: True)
    batch = adapter.prepare_batch(
        [
            DiffusionPagedAttentionRow(
@@ -450,6 +538,48 @@ def test_omni_paged_backend_runs_hunyuan_piecewise_segments(monkeypatch: pytest.
    ]
 def test_piecewise_metadata_snapshots_reused_native_scheduler_buffer(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch)
    shared_scheduler_metadata = torch.zeros(1, dtype=torch.int32)
    build_id = 0
    def build_metadata(**kwargs):
        nonlocal build_id
        shared_scheduler_metadata.fill_(build_id)
        metadata = SimpleNamespace(
            build_id=build_id,
            causal=kwargs["causal"],
            seq_lens=kwargs["seq_lens"].clone(),
            query_start_loc_cpu=kwargs["query_start_loc_cpu"].clone(),
            positions=kwargs["positions"].clone(),
            slot_mappings=kwargs["slot_mappings"].clone(),
            scheduler_metadata=shared_scheduler_metadata,
            prefix_scheduler_metadata=None,
        )
        build_id += 1
        return {"layer-0": metadata}
    monkeypatch.setattr(adapter_module.current_omni_platform, "build_diffusion_kv_attn_metadata", build_metadata)
    batch = adapter.prepare_batch(
        [DiffusionPagedAttentionRow(request_id="req-0", sequence_id=0, query_len=4, seq_len=4)]
    )
    qkv = torch.randn(1, 4, 2, 4)
    metadata = SimpleNamespace(full_attn_spans=[[(1, 3)]], extra={})
    with adapter.activate(batch):
        context = adapter.prepare_layer_context(
            "layer-0",
            qkv,
            qkv,
            qkv,
            omni_attn_metadata=metadata,
        )
    segment_schedules = [item.scheduler_metadata for item in context.piecewise_native_metadata]
    assert [item.item() for item in segment_schedules] == [1, 2, 3]
    assert len({item.data_ptr() for item in segment_schedules}) == 3
 def test_omni_paged_backend_treats_empty_full_spans_as_causal(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, layer, events = _make_adapter(monkeypatch)
    batch = adapter.prepare_batch(
@@ -851,6 +981,28 @@ def test_forward_rejects_batched_layout_that_does_not_match_rows(monkeypatch: py
        adapter.prepare_layer_context("layer-0", qkv, qkv, qkv)
 def test_forward_packs_and_restores_right_padded_heterogeneous_rows(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch)
    batch = adapter.prepare_batch(
        [
            DiffusionPagedAttentionRow(request_id="req-0", sequence_id=0, query_len=1, seq_len=1),
            DiffusionPagedAttentionRow(request_id="req-1", sequence_id=1, query_len=3, seq_len=3),
        ]
    )
    qkv = torch.arange(2 * 3 * 2 * 4, dtype=torch.float32).reshape(2, 3, 2, 4)
    with adapter.activate(batch):
        context = adapter.prepare_layer_context("layer-0", qkv, qkv, qkv)
    assert context.output_scatter_indices.tolist() == [0, 3, 4, 5]
    torch.testing.assert_close(context.query, qkv.reshape(-1, 2, 4)[[0, 3, 4, 5]])
    restored = context.restore_output(context.query)
    assert restored.shape == qkv.shape
    torch.testing.assert_close(restored[0, 0], qkv[0, 0])
    assert torch.count_nonzero(restored[0, 1:]) == 0
    torch.testing.assert_close(restored[1], qkv[1])
 def test_forward_rejects_padded_full_kv(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch)
    batch = adapter.prepare_batch(
@@ -878,6 +1030,41 @@ def test_forward_rejects_padded_full_kv(monkeypatch: pytest.MonkeyPatch) -> None
        adapter.prepare_layer_context("layer-0", query, full_kv, full_kv)
 def test_runner_runtime_switches_from_prefill_to_reusable_denoise_metadata(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch)
    metadata = DiffusionPagedAttentionMetadata(
        prefill_rows=(DiffusionPagedAttentionRow(request_id="req-0", sequence_id=0, query_len=6, seq_len=6),),
        denoise_rows=(
            DiffusionPagedAttentionRow(
                request_id="req-0",
                sequence_id=0,
                query_len=2,
                seq_len=5,
                kv_start_pos=3,
            ),
        ),
    )
    runtime = DiffusionPagedAttentionRuntime(adapter, metadata)
    with set_forward_context(paged_kv_runtime=runtime), runtime.activate():
        set_forward_context_denoise_step_idx(0)
        assert Attention._active_paged_kv_adapter() is adapter
        assert adapter._active_batch.rows == metadata.prefill_rows
        set_forward_context_denoise_step_idx(1)
        assert Attention._active_paged_kv_adapter() is adapter
        denoise_batch = adapter._active_batch
        assert denoise_batch.rows == metadata.denoise_rows
        set_forward_context_denoise_step_idx(2)
        assert Attention._active_paged_kv_adapter() is adapter
        assert adapter._active_batch is denoise_batch
    assert adapter._active_batch is None
 def test_causal_batch_rejects_non_suffix_write(monkeypatch: pytest.MonkeyPatch) -> None:
    adapter, _, _, _ = _make_adapter(monkeypatch, non_causal=False)
    row = DiffusionPagedAttentionRow(
@@ -992,7 +1179,8 @@ def test_layer_adapter_accepts_platform_native_backend_and_uses_rank_local_heads
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
    selected_backends = []
-    impl_cls = Mock(return_value=SimpleNamespace(forward=Mock()))
+    specialized_backends = []
    impl_cls = Mock(return_value=SimpleNamespace(forward=Mock(), do_kv_cache_update=Mock()))
    native_backend = SimpleNamespace(
        get_name=lambda: "ASCEND",
        indexes_kv_by_block_stride=lambda: True,
@@ -1003,16 +1191,28 @@ def test_layer_adapter_accepts_platform_native_backend_and_uses_rank_local_heads
    original_backend_per_kind = {"full": object()}
    config = SimpleNamespace(
        attention_config=SimpleNamespace(backend=None, backend_per_kind=original_backend_per_kind),
        parallel_config=SimpleNamespace(prefill_context_parallel_size=2),
        model_config=SimpleNamespace(dtype=torch.float16),
        cache_config=SimpleNamespace(cache_dtype="auto"),
    )
    def select_backend(**_kwargs):
-        selected_backends.append((config.attention_config.backend, config.attention_config.backend_per_kind))
+        selected_backends.append(
            (
                config.attention_config.backend,
                config.attention_config.backend_per_kind,
                config.parallel_config.prefill_context_parallel_size,
            )
        )
        return native_backend
    monkeypatch.setattr(adapter_module, "get_attn_backend", select_backend)
    monkeypatch.setattr(adapter_module, "set_current_vllm_config", lambda _config: nullcontext())
    monkeypatch.setattr(
        adapter_module.current_omni_platform,
        "get_diffusion_paged_kv_attn_backend",
        lambda backend, *, ulysses_degree: specialized_backends.append((backend, ulysses_degree)) or backend,
    )
    layer = SimpleNamespace(num_heads=8, softmax_scale=0.125)
    spec = FullAttentionSpec(
        block_size=16,
@@ -1031,9 +1231,11 @@ def test_layer_adapter_accepts_platform_native_backend_and_uses_rank_local_heads
        ulysses_degree=2,
    )
-    assert selected_backends == [(adapter_module.AttentionBackendEnum.FLASH_ATTN, {})]
+    assert selected_backends == [(adapter_module.AttentionBackendEnum.FLASH_ATTN, {}, 2)]
    assert specialized_backends == [(native_backend, 2)]
    assert config.attention_config.backend is None
    assert config.attention_config.backend_per_kind is original_backend_per_kind
    assert config.parallel_config.prefill_context_parallel_size == 2
    assert native_layer.num_heads == 4
    assert native_layer.num_kv_heads == 2
    assert native_layer.spec.num_kv_heads == 2
@@ -1094,6 +1296,51 @@ def test_omni_attention_wraps_paged_kernel_with_sp_hooks() -> None:
    assert torch.equal(output, torch.full_like(original, 6))
 def test_omni_attention_strips_paged_ulysses_padding_around_kernel() -> None:
    events: list[tuple[str, tuple[int, ...]]] = []
    class Strategy:
        name = "ulysses"
        def pre_attention(self, query, key, value, metadata):
            return query, key, value, metadata, SimpleNamespace(joint_len=0, joint_strategy="front")
        def post_attention(self, output, _ctx):
            events.append(("post", tuple(output.shape)))
            return output
    class Adapter:
        def prepare_layer_context(self, _layer_name, query, key, value, *, omni_attn_metadata):
            assert omni_attn_metadata is None
            events.append(("kernel_input", tuple(query.shape)))
            assert query[:, -1].tolist() == [[[3.0]]]
            assert torch.equal(query, key)
            assert torch.equal(query, value)
            return SimpleNamespace(query=query)
    class Backend:
        def forward_paged(self, context):
            return context.query
    layer = Attention.__new__(Attention)
    nn.Module.__init__(layer)
    layer.prefix = "layer-0"
    layer.paged_kv_cache_role = "primary"
    layer.attn_backend = SimpleNamespace(supports_paged_kv=True, get_name=lambda: "FLASH_ATTN")
    layer.attention = Backend()
    layer.use_ring = False
    layer._no_parallel_strategy = object()
    layer._get_active_parallel_strategy = lambda: Strategy()
    tensor = torch.arange(5, dtype=torch.float32).reshape(1, 5, 1, 1)
    with set_forward_context(), override_paged_kv_adapter(Adapter()):
        get_forward_context().sp_padding_size = 1
        output = layer._forward_impl(tensor, tensor, tensor)
    assert events == [("kernel_input", (1, 4, 1, 1)), ("post", (1, 5, 1, 1))]
    assert output.flatten().tolist() == [0.0, 1.0, 2.0, 3.0, 0.0]
 def test_omni_attention_rejects_paged_request_for_backend_without_paged_support() -> None:
    class Adapter:
        def prepare_layer_context(self, *args, **kwargs):
@@ -1145,3 +1392,15 @@ def test_omni_attention_keeps_dense_kernel_without_active_adapter() -> None:
    assert events == ["pre", "dense", "post"]
    assert output is qkv
 def test_omni_attention_rejects_scheduler_paged_forward_without_active_adapter() -> None:
    layer = Attention.__new__(Attention)
    nn.Module.__init__(layer)
    layer._scheduler_paged_kv = True
    layer.paged_kv_cache_role = "primary"
    layer._get_active_parallel_strategy = lambda: object()
    qkv = torch.zeros(1, 2, 2, 4)
    with set_forward_context(), pytest.raises(RuntimeError, match="without an active Worker adapter"):
        layer._forward_impl(qkv, qkv, qkv)
@@ -154,8 +154,17 @@ def test_adapter_executes_native_paged_attention_on_non_contiguous_blocks() -> N
    torch.manual_seed(1)
    query = torch.randn(19, _NUM_HEADS, _HEAD_SIZE, dtype=torch.float16, device=device)
-    key = torch.randn_like(query)
+    packed_qkv = torch.randn(
-    value = torch.randn_like(query)
+        19,
        3 * _NUM_HEADS * _HEAD_SIZE,
        dtype=torch.float16,
        device=device,
    )
    _, key_flat, value_flat = packed_qkv.split(_NUM_HEADS * _HEAD_SIZE, dim=-1)
    key = key_flat.reshape(19, _NUM_HEADS, _HEAD_SIZE)
    value = value_flat.reshape(19, _NUM_HEADS, _HEAD_SIZE)
    assert not key.is_contiguous()
    assert not value.is_contiguous()
    omni_backend = FlashAttentionImpl(
        num_heads=_NUM_HEADS,
        head_size=_HEAD_SIZE,
Author	SHA1	Message	Date
hsliu_ustc	7d41ea6183	Merge remote-tracking branch 'origin/main' into codex-api-server-user-guide Signed-off-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> # Conflicts: # docs/user_guide/examples/online_serving/text_to_audio.md # examples/online_serving/stable_audio/README.md	2026-09-01 11:23:08 +08:00
Alex Brooks	a65e89cbd4	[Core] Validate PipelineConfig in `__post_init__` (#6829 ) Signed-off-by: Alex Brooks <albrooks@redhat.com>	2026-09-01 11:17:57 +08:00
Yuxuan RhoandHongsheng Liu	7992c2213e	[Config] Remove diffusion_batch_size and use max_num_seqs instead. (#6484 ) Signed-off-by: LyxWxj <yuxuanlou621@foxmail.com> Signed-off-by: Yuxuan Rho <104440472+LyxWxj@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-09-01 01:02:54 +00:00
XRay2016	8e26dce1f4	[Diffusion] Add Boogu Edit-Turbo support docs and tests (#6701 ) Signed-off-by: xRay2016 <1150722393@qq.com>	2026-09-01 08:41:23 +08:00
Inesh Reddy ChappidiandGao Han	eb11446b7f	[Bugfix][Core] Do not advance the segment watermark past an unmaintained request counter (#6834 ) Signed-off-by: Inesh Reddy Chappidi <ineshreddy249@gmail.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-31 15:27:13 +00:00
Gao Han	705e781b0e	[Bugfix][MiniCPM-o] Checkpoint playback before final input commit (#6821 ) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-31 15:18:29 +00:00
Gao Han	b81aeb7b86	[CI][MiniCPM-o] Make audio consistency checks deterministic (#6828 ) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-31 13:08:49 +00:00
wangyu	3d1ad0c0f4	[CI][Bugfix]Drop dummy-weight Qwen3-TTS Base from Ready CI (#6861 ) Signed-off-by: wangyu <410167048@qq.com>	2026-08-31 12:13:26 +00:00
Wang fuyin zwhzzz0821 zwhzzz0821	fe4a2a7477	[Diffusion] Add paged KV cache support for HunyuanImage3 DiT on NPU and GPU (#6563 ) Signed-off-by: Acerak01-fy <wfy2003324@163.com> Signed-off-by: zwhzzz0821 <zwhzzz0821@users.noreply.github.com> Signed-off-by: zwhzzz0821 <2831474076@qq.com> Co-authored-by: zwhzzz0821 <zwhzzz0821@users.noreply.github.com> Co-authored-by: zwhzzz0821 <2831474076@qq.com>	2026-08-31 10:15:46 +00:00
Bo Li	6f16f9546d	[Bugfix][Kernel] Stage strided Ulysses QKV with copy engine (#6814 ) Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com>	2026-08-31 17:28:49 +08:00
Alex Brooks	58cb8de68b	[CI] Fix Speech CI Validation Messages (#6827 ) Signed-off-by: Alex Brooks <albrooks@redhat.com>	2026-08-31 08:13:59 +00:00
Rakesh Kariya	b864374089	[Bugfix][MiMo-Audio] Restore make_empty_intermediate_tensors on the talker for vLLM 0.28 (#6803 ) Signed-off-by: rk9595 <rakesh.kariya@somaiya.edu>	2026-08-31 15:26:35 +08:00
WeiQing Chen	2052deef07	[Bugfix][Diffusion] Load plugins in spawned stage process (#6750 ) Signed-off-by: david6666666 <530634352@qq.com>	2026-08-31 14:32:07 +08:00
汪志鹏andHongsheng Liu	86b85c078b	[Feature][MiniMax-H3] Fuse the FastH3 four-step adapter at load time (#6714 ) Signed-off-by: princepride <wangzhipeng628@gmail.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-31 12:04:58 +08:00
Inesh Reddy Chappidi	5e06d2e73c	[CI/Build] Align duplex stage-input deadline with the e2e client timeout (#6831 ) Signed-off-by: Inesh Reddy Chappidi <ineshreddy249@gmail.com>	2026-08-31 02:55:36 +00:00
akshatvishu	f321cdbe16	[Bugfix] Use Seed-TTS reference audio with MiniCPM-o 4.5 (#6628 ) Signed-off-by: akshatvishu <akshatnayak197@gmail.com>	2026-08-31 10:27:34 +08:00
WeiQing Chen	4cca28e9bd	[Doc] Update vLLM-Omni WeChat QR code (#6832 ) Signed-off-by: David Chen <530634352@qq.com>	2026-08-31 09:44:33 +08:00
TJian	ded8934626	[AMD] [CI] Fix tests for v0.28.0 (#6830 ) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>	2026-08-31 09:10:27 +08:00
SYLAR	759aa4ffeb	[Diffusion][Performance] Optimize MiniMax-H3 video output transfer (#6824 ) Signed-off-by: lishunyang12 <lishunyang12@163.com>	2026-08-31 07:07:34 +08:00
mokekeandmokashliu	1a9d598d2b	[Frontend] Encode interleaved video frames in parallel (#6776 ) Signed-off-by: mokashliu <mokashliu@tencent.com> Co-authored-by: mokashliu <mokashliu@tencent.com>	2026-08-31 07:04:27 +08:00
m0g3r	8fe913e461	[Core] Flag pipelines with no terminal output stage in PipelineConfig.validate() (#6291 ) Signed-off-by: m0g3r <87276771+m0g3r@users.noreply.github.com>	2026-08-30 21:33:06 +00:00
Gao Han	71748388f8	[Bugfix] Recover Omni benchmark TPOT from Stage 0 metrics (#6818 ) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-31 00:43:25 +08:00
Gao Han	f7f48929db	[Bugfix][Diffusion] Restore inline single-stage execution (#6813 ) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-30 21:40:47 +08:00
Weiming Liao Hongsheng Liu amy-why-3459	18f01befa7	[NPU] upgrade to v0.28.0 (#6674 ) Signed-off-by: Weiming Liao <liaowm5@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-30 12:28:35 +00:00
Yueqian Lin	1d75e63f7f	[CI/Build] Add dots.tts weekly e2e coverage (#6174 ) Signed-off-by: Yueqian Lin <linyueqian@outlook.com>	2026-08-30 11:37:08 +00:00
akshatvishu	c41925689d	[BugFix][Higgs-Audio-V3] Fix transcript and concurrent sampling tests (#6422 ) Signed-off-by: akshatvishu <akshatnayak197@gmail.com>	2026-08-30 08:52:15 +00:00
Yueqian Lin	a6b559cf54	[Bugfix][MiniMax-Music3] Resolve stage subdirs against a real snapshot (#6640 ) Signed-off-by: Yueqian Lin <linyueqian@outlook.com>	2026-08-30 06:51:24 +00:00
baonudesifeizhai	8be601ed7f	[Diffusion] Add native SymmMem Fast Ulysses transport (#6340 ) Signed-off-by: roG0d <baonudesifeizhai@gmail.com>	2026-08-30 04:14:21 +00:00
y-null	99a2ae1fc8	[Bugfix][MiniCPM-o] Anchor duplex soft-interrupt listen sandwich before commit (#6772 ) Signed-off-by: y_null <y_null@qq.com>	2026-08-30 03:25:43 +00:00
Sy03 Nathan Price Eric	21dd4b33ac	[Bugfix][MiniCPM-o] Stabilize native duplex streaming and Stage-1 handoffs (#6529 ) Signed-off-by: Sy03 <1370724210@qq.com> Co-authored-by: Nathan Price <nathan@abridge.com> Co-authored-by: Eric(Qun) <186020953+0xtoward@users.noreply.github.com>	2026-08-29 13:05:24 -07:00
WeiQing ChenandGao Han	9a88d2d21d	[diffusion][bugfix] Separate Diffusers hooks from model metadata (#6749 ) Signed-off-by: david6666666 <530634352@qq.com> Signed-off-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-29 17:37:14 +00:00
YancyandAsthenia	73893e73ea	[Bugfix] Fix create_error_response import across vLLM versions (#6773 ) Signed-off-by: Asthenia <asthenia0412@gmail.com> Co-authored-by: Asthenia <asthenia0412@gmail.com>	2026-08-29 17:01:51 +00:00
NATURE	87736a3cc5	[Bugfix][MiniCPM-o] Bound full-duplex Talker context with sliding recompute (#6626 ) Signed-off-by: natureofnature <wzliu@connect.hku.hk>	2026-08-30 00:49:42 +08:00
Sy03	3adee400a0	[Bugfix][Qwen3-TTS] Handle codec generations that exhaust their token budget (#6728 ) Signed-off-by: Sy03 <1370724210@qq.com>	2026-08-29 08:59:05 -07:00
Gao Han	676a35fa8b	Fix MiniCPM-o concurrent audio first-packet latency (#6767 ) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-29 23:27:47 +08:00
Honghan Zhuandamy-why-3459	72a1ce488b	[Bugfix] Avoid invalid TPOT metrics (#6696 ) Signed-off-by: zhuhh97 <zhuhonghan@huawei.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-29 11:56:37 +00:00
wangyu	e338f0ece3	[CI]Split GPU jobs by cards_* and reject hand-written SKU marks (#6650 ) Signed-off-by: wangyu <410167048@qq.com> Signed-off-by: [Your Name] <your.email@example.com>	2026-08-29 18:42:33 +08:00
HAAZZZEEEE	c50b7de167	[NPU][Diffusion] MiniMax-H3 RainFusion end_step tail fallback (#6037 ) Signed-off-by: HAAZZZEEEE <wang-taicheng@qq.com>	2026-08-29 18:36:05 +08:00
psv666	a5b8c82788	[Bugfix][MiniCPM-o] Fix duplex camera frame fixture (#6757 ) Signed-off-by: psv666 <2693925048@qq.com>	2026-08-29 09:57:53 +00:00
ooooooye	5a5bc55c25	[Feature][Diffusion] Allow online INT8 quantization with DLO AllGather (#6573 ) Signed-off-by: brandneway <gyuan4892@gmail.com>	2026-08-29 09:50:30 +00:00
Wenjia LiandWeiQing Chen	d9980c75e8	Perf/minimax h3 dit swiglu rope cache npu (#6410 ) Signed-off-by: Wenjia Li <wjialish@gmail.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-29 09:14:52 +00:00
Nguyen Kha Nhat Long Cursor Samit	bb5bf8fd0a	[Bugfix] Fix colocate-async sleep admission race and deliver AR abort tokens (#6367 ) Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com> Signed-off-by: Cursor <cursoragent@cursor.com> Signed-off-by: Samit <285365963@qq.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Samit <285365963@qq.com>	2026-08-29 09:11:39 +00:00
psv666andGao Han	5c47fd0bc8	[Bugfix][Qwen3-Omni] Handle missing packed modules mapping (#6748 ) Signed-off-by: psv666 <2693925048@qq.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>	2026-08-29 15:24:49 +08:00
Zijun HuangandWeiQing Chen	08cdee8dc4	[Feature] Support MiniMax-H3 FlashGen native LoRA with the legacy manager (#6666 ) Signed-off-by: Zijun Huang <zijunhua@usc.edu> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-29 06:52:19 +00:00
abrahamzewoudie Hongsheng Liu amy-why-3459	143ccc5a2f	Fix Qwen3-Omni audio encoder TP when heads are not divisible by TP size (#4322 ) Signed-off-by: abrahamzewoudie <azewoudi@amd.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-29 14:52:02 +08:00
WeiQing Chen	55a226dcf1	[video][bugfix] Accept semantic video output in offline examples (#6747 ) Signed-off-by: david6666666 <530634352@qq.com>	2026-08-29 06:26:28 +00:00
Jiatai Wang	f66f05633c	[Bugfix] Honor request seed in MOSS-TTS adapters (#6543 ) Signed-off-by: JiataiWang <wangjiatai@proton.me>	2026-08-28 22:44:13 -07:00
ZacheryAUandamy-why-3459	457c6e2793	[bugfix][MiniCPM-o] Re-land empty full-payload and async_chunk=false processor tests (#6745 ) Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-29 13:39:54 +08:00
boatman	bfe8967ca2	[Bugfix][StepAudio2] Fix async-chunk metadata mismatch and boundary glitches (#5917 ) Signed-off-by: boatman <109857087+sphinxkkkbc@users.noreply.github.com>	2026-08-28 21:45:24 -07:00
WeiQing ChenandOpenAI Codex	fbc93a0e08	[ci][bugfix] Fix MiniMax-H3 FP8 quality test OOM (#6742 ) Signed-off-by: david6666666 <530634352@qq.com> Co-authored-by: OpenAI Codex <codex@openai.com>	2026-08-29 12:11:31 +08:00
Bob Tian	d934a5f7b8	[Bugfix] Report engine-queued requests as waiting in vllm_omni:num_requests gauges (#6549 ) Signed-off-by: Bob Tian <bobzetian@google.com>	2026-08-29 03:58:26 +00:00
Inesh Reddy Chappidiandamy-why-3459	bbb3ae01a1	[CI] Gate qwen3-omni no-async-chunk perf on mean_audio_rtf (#6743 ) Signed-off-by: Inesh Reddy <ineshreddy249@gmail.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-28 20:28:58 -07:00
Zhang BowenandHongsheng Liu	46cfa8435b	[Model][Frontend] MiniMax-H3: Parallelize MP4 response frame conversion (#6499 ) Signed-off-by: MosCloud <bwzhang1991@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-28 21:22:00 -06:00
fan2956andWeiQing Chen	0dc8f45906	[Feat] Minimax H3 support ref and tgt video sparse attention on NPU (#6518 ) Signed-off-by: fan2956 <zhoufan53@huawei.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-29 11:10:53 +08:00
Allen Wu TRAE CLI TRAE CLI	d94061201d	[Bugfix][Engine] Make abort idempotent during shutdown (#6327 ) Signed-off-by: Allen Wu <allenwu2795@gmail.com> Co-authored-by: TRAE CLI <noreply@bytedance.com> Co-authored-by: TRAE CLI <traecli@bytedance.com>	2026-08-28 19:51:25 -07:00
Mu GuanLinandWeiQing Chen	9558fd2ea8	[Bugfix] Scope MiniMax-H3 cuDNN SDPA state (#6710 ) Signed-off-by: mglyn <1203789601@qq.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-28 19:40:50 -07:00
Mu GuanLin	55a8bd6b80	[CI][Bugfix] Align the LTX-2 FP8 quality gate with the default recipe (#5831 ) Signed-off-by: mglyn <1203789601@qq.com>	2026-08-29 02:24:09 +00:00
Honghan Zhu	f584729f24	[BugFix] Fix local media access for diffusion speech (#6622 ) Signed-off-by: zhuhh97 <zhuhonghan@huawei.com>	2026-08-29 02:01:24 +00:00
Chendi.Xue	034af1f657	[CI][XPU]switch to use vLLM base docker for XPU (#6727 ) Signed-off-by: Chendi Xue <chendi.xue@intel.com>	2026-08-29 09:03:00 +08:00
Hongsheng Liuandhsliuustc0106	8bfad23f98	[Core] Restore topology-aware diffusion worker titles (#6722 ) Signed-off-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com> Co-authored-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>	2026-08-29 08:45:57 +08:00
Gao Han	6b9ce05feb	Revert "[bugfix][MiniCPM-o] Fix offline_inference/test_minicpmo_4_5.py and online_serving/ one" (#6730 )	2026-08-29 08:14:02 +08:00
ZacheryAUandamy-why-3459	6d84409a68	[bugfix][MiniCPM-o] Fix offline_inference/test_minicpmo_4_5.py and online_serving/ one (#5464 ) Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-28 19:20:09 +00:00
Khairul Kabir khairulkabir1661 Claude Opus 4.6	ebb3b20576	[Diffusion] Make Flux2Pipeline text_encoder_out_layers configurable (#6390 ) Co-authored-by: khairulkabir1661 <khairulkabir1661@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>	2026-08-28 09:05:41 -06:00
Wallbreazzz Canlin Guo Canlin Guo	be335a86f9	fix(npu): extend bf16 autocast to NPU in MossAudioTokenizer decode (#6664 ) Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com> Signed-off-by: Canlin Guo <961750412@qq.com> Co-authored-by: Canlin Guo <961750412@qq.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com>	2026-08-28 04:01:49 -07:00
Hongsheng Liu	d5aced9296	Revert "[Core] Make diffusion worker titles topology-aware" (#6717 )	2026-08-28 17:30:03 +08:00
Wallbreazzz	4ad455e61e	[Feature][TTS] adaptive chunk ramp (Phase 2 buffer-feedback controller) (#6001 ) Signed-off-by: Wallbreazzz <110282866+Wallbreazzz@users.noreply.github.com>	2026-08-28 02:11:05 -07:00
amy-why-3459	da94c3ebe1	[Bugfix][MiniCPM-o] Bind omni-duplex video frames to the unit they close (#6404 ) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-28 16:58:20 +08:00
prettygirlisnotme	13173d6bbb	[Core] Make diffusion worker titles topology-aware (#6307 ) Signed-off-by: Dragonborn <1078757180@qq.com>	2026-08-28 16:00:30 +08:00
Sy03	009b992c5b	[Bugfix][MiniCPM-o] Isolate duplex handoff failures and barge-in (#6170 ) Signed-off-by: Sy03 <1370724210@qq.com>	2026-08-28 15:46:41 +08:00
Mu GuanLinandHongsheng Liu	04b97a5922	[Perf] MiniMax-H3 VAE Decoder Ops (#6607 ) Signed-off-by: mglyn <1203789601@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-28 15:14:33 +08:00
TJianandroot	15c6ab8fa7	[AMD][CI] Fix AMD CI (Partial) (#6704 ) Signed-off-by: root <root@tw020.mia.tensorwave.lan> Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Co-authored-by: root <root@tw020.mia.tensorwave.lan>	2026-08-28 14:34:11 +08:00
zzh	c201fd77c6	Add HYvideo1.5 benchmark (#6349 ) Signed-off-by: zzh <943967662@qq.com>	2026-08-28 14:31:03 +08:00
Canlin Guo lishunyang12 yuanwu	b6a4baa659	[Feature] Support text encoder Disaggregation for H3 (#5885 ) Signed-off-by: Canlin Guo <canlinguosdu@gmail.com> Signed-off-by: lishunyang12 <lishunyang12@163.com> Signed-off-by: yuanwu <yuan.wu@intel.com> Co-authored-by: lishunyang12 <lishunyang12@163.com> Co-authored-by: yuanwu <yuan.wu@intel.com>	2026-08-28 14:27:05 +08:00
WeiQing Chen	ed837e350e	[Test] Stabilize MiniMax-H3 reference accuracy compile (#6688 ) Signed-off-by: david6666666 <530634352@qq.com>	2026-08-28 14:18:28 +08:00
BruceLoveDecimalandamy-why-3459	ab8d45d3cc	[Bugfix][MiniCPM-o] Retire CFM DiT CUDA graphs a generation at a time (#6587 ) Signed-off-by: BruceLoveDecimal <liuqihao970610@gmail.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-28 14:07:57 +08:00
Eugene Zhao	f8f8aa8eb7	[Bugfix] Honor stage selection in online profiler endpoints (#6609 ) Signed-off-by: qujing226 <qujing226@gmail.com>	2026-08-28 11:41:47 +08:00
Hongsheng Liu	6ec5ab11bd	[Bugfix] Make HWR filesystem lifecycle transitions fail-closed (#6692 ) Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-08-28 09:30:17 +08:00
Hongsheng Liuandhsliuustc0106	4245fd6955	[Fix] Restore MammothModa2 support (#6694 ) Signed-off-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com> Co-authored-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>	2026-08-28 08:32:32 +08:00
Mu GuanLinandHongsheng Liu	5e4daaf197	[Diffusion] Add LTX-2.5 Diffusion VAE decoder support (#6189 ) Signed-off-by: mglyn <1203789601@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-27 23:54:40 +08:00
NATUREandamy-why-3459	262ff14693	[Bugfix][MiniCPM-o] Bound native auto-response continuation (#6630 ) Signed-off-by: natureofnature <wzliu@connect.hku.hk> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-27 19:51:47 +08:00
Deep Shahandamy-why-3459	069e6c5fbc	[CI/Build][MiniCPM-o] Derive duplex admission-probe limit from the deploy config (#6678 ) Signed-off-by: Deep Shah <deep@socratic.co> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-08-27 19:20:01 +08:00
NATURE	54c1fc64b7	[Bugfix] Gate async-chunk segment resume by session mode (#6680 ) Signed-off-by: natureofnature <wzliu@connect.hku.hk>	2026-08-27 18:52:09 +08:00
Hongsheng Liu	a79dcdee5c	docs: address API guide review feedback Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-17 13:31:36 +08:00
Hongsheng Liu	159ac2161a	docs: standardize API example ports Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-13 23:34:28 +08:00
Hongsheng Liu	8327f26216	docs: document realtime and specialized API servers Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-13 21:46:24 +08:00
Hongsheng Liu	5c6ed9b4a9	docs: add API server endpoint guide Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-13 21:28:36 +08:00