Merge branch 'main' into fix/pr4709-review-followup

[bugfix] Use FlowMatchEulerDiscreteScheduler from diffusers 0.39.0 with correct RNG handling for upcoming Cosmos3 distilled models (#5176 )
Signed-off-by: Bartosz Stefaniak <bstefaniak@nvidia.com>
2026-07-18 11:28:29 +08:00 · 2026-07-17 11:38:04 -07:00 · 2026-07-17 17:32:48 +08:00 · 2026-07-17 03:26:35 +00:00 · 2026-07-16 15:01:21 -07:00 · 2026-07-16 12:33:57 -07:00
278 changed files with 16322 additions and 4451 deletions
@@ -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: "v0.24.0"
+          VLLM_VERSION: "v0.25.0"
        priority: 100
        timeout_in_minutes: 60
        soft_fail: false
@@ -78,8 +78,89 @@ steps:
      echo "--- Image pushed successfully"
      echo "$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"

+  - label: ":buildkit: Build and Push NPU Test Image - A3"
+    key: image-build-a3
+    agents:
+      queue: "ascend-a2b3"
+      resource_class: "npu-1"
+    plugins:
+      - kubernetes:
+          metadata:
+            annotations:
+              vault.hashicorp.com/agent-init-first: "true"
+              vault.hashicorp.com/agent-inject: "true"
+              vault.hashicorp.com/agent-inject-perms-ca.pem: "0400"
+              vault.hashicorp.com/agent-inject-perms-cert.pem: "0400"
+              vault.hashicorp.com/agent-inject-perms-config.json: "0400"
+              vault.hashicorp.com/agent-inject-perms-key.pem: "0400"
+              vault.hashicorp.com/agent-inject-secret-ca.pem: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-secret-cert.pem: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-secret-config.json: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-secret-key.pem: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-template-ca.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.RootCA }}\n{{- end }}"
+              vault.hashicorp.com/agent-inject-template-cert.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaCert }}\n{{- end }}"
+              vault.hashicorp.com/agent-inject-template-config.json: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.dockerConfig }}\n{{- end }}"
+              vault.hashicorp.com/agent-inject-template-key.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaKey }}\n{{- end }}"
+              vault.hashicorp.com/agent-pre-populate-only: "true"
+              vault.hashicorp.com/agent-run-as-group: "1000"
+              vault.hashicorp.com/agent-run-as-user: "1000"
+              vault.hashicorp.com/agent-service-account-token-volume-name: token-vol
+              vault.hashicorp.com/role: ascend-gha-runners
+              vault.hashicorp.com/secret-volume-path: /home/user/.docker/
+              vault.hashicorp.com/tls-skip-verify: "true"
+          podSpecPatch:
+            volumes:
+            - name: token-vol
+              projected:
+                defaultMode: 420
+                sources:
+                - serviceAccountToken:
+                    audience: api
+                    expirationSeconds: 600
+                    path: token
+    env:
+      VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}-a3"
+      IMAGE_NAME: "vllm-omni-ci-npu"
+      IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
+      BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
+    command: |
+      set -ex
+
+      echo "--- Building and pushing NPU Test Image"
+      echo "Image: $${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+      echo "buildkitd address: $${BUILDKITD_ADDR}"
+
+      if ! command -v buildctl &> /dev/null; then
+        echo "Installing buildctl..."
+        mkdir -p /tmp/buildkit
+        BUILDKIT_VERSION="v0.29.0"
+        wget -q "https://github.com/moby/buildkit/releases/download/$${BUILDKIT_VERSION}/buildkit-$${BUILDKIT_VERSION}.linux-arm64.tar.gz" -O /tmp/buildkit.tar.gz
+        tar -xzf /tmp/buildkit.tar.gz -C /tmp/buildkit
+        cp /tmp/buildkit/bin/buildctl /usr/local/bin/
+      fi
+
+      export DOCKER_CONFIG=/home/user/.docker
+      buildctl \
+        --addr="$${BUILDKITD_ADDR}" \
+        --tlscacert=/home/user/.docker/ca.pem \
+        --tlscert=/home/user/.docker/cert.pem \
+        --tlskey=/home/user/.docker/key.pem \
+        build \
+        --frontend dockerfile.v0 \
+        --local context=. \
+        --local dockerfile=./docker \
+        --opt filename=Dockerfile.npu.ci.a3 \
+        --secret id=dockerconfig,src=/home/user/.docker/config.json \
+        --output type=image,name=$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG},push=true \
+        --progress=plain
+
+      echo "--- Image pushed successfully"
+      echo "$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+
  - label: "Upload Ready Pipeline"
-    depends_on: image-build
+    depends_on:
+      - image-build
+      - image-build-a3
    key: upload-ready-pipeline
    if: >-
      build.branch != "main" && (
@@ -94,7 +175,9 @@ steps:

  # L4 Test — main+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
  - label: "Upload Nightly Pipeline"
-    depends_on: image-build
+    depends_on:
+      - image-build
+      - image-build-a3
    key: upload-nightly-pipeline
    if: >-
      (build.branch == "main" && build.env("NIGHTLY") == "1") ||
@@ -13,12 +13,33 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(
 )"

 cd "${omni_source_dir}"
+
+# The XPU base image is ~37GB; the default gzip layer exporter is single-threaded
+# and takes ~16min to compress it. zstd is multi-threaded (uses all cores) at a
+# similar ratio; set EXPORT_COMPRESSION=uncompressed to skip compression entirely
+# for local-only images. This requires the containerd image store (buildx docker
+# driver), which is the default here.
+EXPORT_COMPRESSION="${EXPORT_COMPRESSION:-zstd}"
+if [ "${EXPORT_COMPRESSION}" = "uncompressed" ]; then
+    export_args=(--output "type=image,name={{IMAGE}},compression=uncompressed")
+else
+    export_args=(--output "type=image,name={{IMAGE}},compression=${EXPORT_COMPRESSION},compression-level=3,force-compression=true")
+fi
+
+docker_build() {
+    # $1 = image name; remaining args passed through to docker build.
+    local image="$1"
+    shift
+    local out=("${export_args[@]/'{{IMAGE}}'/${image}}")
+    docker build "${out[@]}" "$@" -f docker/Dockerfile.xpu .
+}
+
 if [ -z "$(docker images -q "${base_image_name}")" ]; then
-    docker build --target vllm-base -t "${base_image_name}" --build-arg "VLLM_VERSION=${VLLM_VERSION}" -f docker/Dockerfile.xpu .
+    docker_build "${base_image_name}" --target vllm-base --build-arg "VLLM_VERSION=${VLLM_VERSION}"
 fi

 # Try building the docker image
-docker build --build-arg "VLLM_BASE=${base_image_name}" --build-arg "VLLM_VERSION=${VLLM_VERSION}" -t "${image_name}" -f docker/Dockerfile.xpu .
+docker_build "${image_name}" --build-arg "VLLM_BASE=${base_image_name}" --build-arg "VLLM_VERSION=${VLLM_VERSION}"

 # Setup cleanup
 remove_docker_container() {
@@ -173,6 +173,9 @@ steps:
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
        commands:
+          # 40m (was 20m): post-rebase runs take 9-12 min warm, but cold-cache
+          # runs exceed 20m and were killed without failure output (a cold run
+          # died at 25.2m wall under the old limit).
          - |
            timeout 40m bash -c "
              export VLLM_LOGGING_LEVEL=DEBUG
@@ -200,6 +203,9 @@ steps:
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
        commands:
+          # 40m (was 20m): post-rebase runs take 9-13 min warm, but cold-cache
+          # runs exceed 20m and were killed without failure output (a cold run
+          # died at 25.2m wall under the old limit).
          - |
            timeout 40m bash -c "
              export VLLM_LOGGING_LEVEL=DEBUG
@@ -262,10 +268,13 @@ steps:
          - vllm_omni/entrypoints/openai/tts_adapters/indextts2.py
          - vllm_omni/config/pipeline_registry.py
          - vllm_omni/model_executor/models/registry.py
-        timeout_in_minutes: 60
+        # 40m/35m (was 20m/20m): post-rebase runs take a steady ~12.5 min warm,
+        # but a cold run was killed at 20.7m under the old limit with no
+        # failure output; the rebase-time raise to 60m/45m was oversized.
+        timeout_in_minutes: 40
        commands:
          - |
-            timeout 45m bash -c "
+            timeout 35m bash -c "
              export VLLM_LOGGING_LEVEL=DEBUG
              pytest -s -v tests/e2e/online_serving/test_indextts2.py tests/e2e/offline_inference/test_indextts2.py -m 'advanced_model and cuda' --run-level 'advanced_model'
            "
@@ -290,6 +299,8 @@ steps:
          - vllm_omni/model_executor/models/qwen3_omni/
          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
          - vllm_omni/deploy/qwen3_omni_moe.yaml
+        # 50m (was 30m): post-rebase runs take 15-21 min warm; 30m left no
+        # headroom for cold-cache runs, which were killed without output.
        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'
@@ -12,6 +12,9 @@ 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"
+        # 120m (was 90m): post-rebase nightlies (#2474/#2486) run 59-71 min —
+        # too close to the old 90m ceiling, which #2444 already hit (killed at
+        # 90.8m with no failure output).
        timeout_in_minutes: 120
        commands:
          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
@@ -153,7 +156,10 @@ steps:

      - label: ":full_moon: Omni · Perf Test · No Async Chunk"
        key: nightly-omni-performance-no-async-chunk
-        timeout_in_minutes: 300
+        # 180m: post-rebase nightlies (#2474/#2486) run 41-44 min; the
+        # rebase-time raise to 300m targeted a pre-rebase 179.4m outlier and
+        # is no longer needed.
+        timeout_in_minutes: 180
        commands:
          - export BENCHMARK_DIR=tests/dfx/perf/results
          - |
@@ -376,7 +382,10 @@ steps:
    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
    steps:
      - label: ":full_moon: TTS · Function Test with L4"
-        timeout_in_minutes: 120
+        # 90m: post-rebase nightlies (#2474/#2486) run ~27 min; the rebase-time
+        # raise to 120m targeted pre-rebase 72-87 min runs and is no longer
+        # needed.
+        timeout_in_minutes: 90
        commands:
          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
          - export VLLM_USE_DEEP_GEMM="0"
@@ -650,6 +659,8 @@ steps:
            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
            tests/e2e/online_serving/test_bagel_expansion.py
+            tests/e2e/online_serving/test_krea2_expansion.py
+            tests/e2e/offline_inference/test_krea2_expansion.py
            -m "full_model and diffusion and H100 and not distributed_cuda"
            --run-level "full_model"
        agents:
@@ -693,6 +704,7 @@ steps:
            pytest -sv
            tests/e2e/online_serving/test_qwen_image_expansion.py
            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            tests/e2e/offline_inference/test_krea2_expansion.py
            -m "full_model and diffusion and H100 and distributed_cuda"
            --run-level "full_model"
        agents:
@@ -858,7 +870,10 @@ steps:


      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
-        timeout_in_minutes: 120
+        # 60m: post-rebase nightlies (#2474/#2486) run ~18 min; the rebase-time
+        # raise to 120m targeted pre-rebase 48-58 min runs and is no longer
+        # needed.
+        timeout_in_minutes: 60
        commands:
          - pytest -sv tests/e2e/ -k "not test_wan and not test_bagel_expansion and not hunyuan" -m "full_model and diffusion and L4" --run-level "full_model" --ignore=tests/e2e/accuracy
        agents:
@@ -1486,9 +1501,12 @@ steps:
            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
            EXIT1=$$?
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
+            EXIT2=$$?
            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
-            exit $$EXIT1
+            if [ $$EXIT1 -ne 0 ]; then exit $$EXIT1; fi
+            exit $$EXIT2
        agents:
          queue: "mithril-h100-pool"
        plugins:
@@ -22,3 +22,31 @@ steps:
          HF_TOKEN: "${HF_TOKEN}"
        commands:
          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k npu_i2v --run-level "full_model"
+  - group: ":card_index_dividers: TTS Model Test"
+    key: nightly-tts-test-group
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
+    steps:
+      - label: ":full_moon: TTS · Perf Test"
+        agents:
+          queue: "ascend-a3"
+          resource_class: "npu-2"
+        image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a3"
+        plugins:
+          - kubernetes:
+              podSpecPatch:
+                imagePullSecrets:
+                  - name: swr-secret
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m npu --test-config-file tests/dfx/perf/tests/test_tts.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
@@ -334,7 +334,7 @@ Classify the model's **CI priority** first:
 | **L1** | `tests/diffusion/models/{slug}/`, `tests/diffusion/cache/`, transformer unit tests | `core_model` + `cpu` | `test-ready.yml` | Weight remap, `_sp_plan`, cache enabler registration, shape contracts |
 | **L2** | `tests/e2e/online_serving/test_{slug}.py` (and offline if the category is offline-first) | **`core_model` + `advanced_model`** (both on baseline smoke) + `diffusion` + `@hardware_test` / `hardware_marks` | `test-ready.yml` | Default deploy smoke — minimal `num_inference_steps`, single prompt |
 | **L3** | `tests/e2e/online_serving/test_{slug}.py` **and** `tests/e2e/offline_inference/test_{slug}.py` when offline matters | Baseline smoke: **`core_model` + `advanced_model`**; heavier cases: `advanced_model` only (+ `diffusion`) | `test-merge.yml` **or** merged into nightly diffusion function job | Real weights, streaming/API paths, LoRA/offload smoke |
-| **L4** | `tests/e2e/online_serving/test_{slug}_expansion.py` (+ offline expansion if needed) | `full_model` + `diffusion` | `test-nightly.yml` (X2I / X2V / X2A function groups) | Feature combos per [#1832](https://github.com/vllm-project/vllm-omni/issues/1832); perf → `tests/dfx/perf/tests/test_{model}_vllm_omni.json` |
+| **L4** | `tests/e2e/online_serving/test_{slug}_expansion.py` (+ offline expansion if needed) | `full_model` + `diffusion` | `test-nightly.yml` (X2I / X2V / X2A function groups) | Feature combos per [#1832](https://github.com/vllm-project/vllm-omni/issues/1832); perf → `tests/dfx/perf/tests/test_{model}_vllm_omni.json` with per-case **`mark`** (`hardware_marks` + `full_model` + `diffusion`) |

 **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`).

@@ -97,7 +97,7 @@ Existing references: `tests/e2e/offline_inference/test_qwen2_5_omni.py` (L2-styl
 | **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `omni_runner_handler`. **Function (isolation only):** `omni_runner_function` + `omni_runner_handler_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
 | **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `openai_client`. **Function (isolation only):** `omni_server_function` + `openai_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
 | **Documentation / runnable examples** | `tests/examples/offline_inference/`, `tests/examples/online_serving/` | **Offline docs (preferred):** extract Python/Bash blocks from the doc README (e.g. `ReadmeSnippet.extract_readme_snippets`), `pytest.mark.parametrize` each snippet, run via `example_runner.run` with a stable `output_subfolder`. **Online docs:** copy client/request scripts into dedicated tests and keep them in sync with the doc page. | Usually **L4**: `advanced_model`, often `example` plus hardware marks matching the nightly docs-example job (see `.buildkite/test-nightly.yml`). Full conventions: [docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md) (introduced in [PR #1910](https://github.com/vllm-project/vllm-omni/pull/1910): naming, output directory layout, skip rules, avoid trimming `num_inference_steps` without a strong CI reason). |
-| **Performance / benchmark** | `tests/dfx/perf/tests/*.json` + `run_*_benchmark.py` | JSON or script-driven server + load config; assert explicit metrics / baselines | L4 Perf: `full_model` + `benchmark`; wire `test-nightly.yml` Perf steps |
+| **Performance / benchmark** | `tests/dfx/perf/tests/*.json` + `run_*_benchmark.py` | JSON or script-driven server + load config; assert explicit metrics / baselines | L4 Perf: JSON `mark` with `full_model` + `omni`/`tts`/`diffusion`; wire `test-nightly.yml` Perf steps |
 | **Invalid parameter / negative HTTP validation** | `tests/dfx/reliability/invalid_param_test/` | Live `omni_server` + low-level `send_*_http_request` with `err_code` / `err_message` | `pytest.mark.slow` + `omni` / `tts` / `diffusion` + `@hardware_marks` (`H100` or `L4`); CI in **`test-weekly.yml`** (not ready/merge/nightly) |

 **If the user’s test plan includes invalid parameter validation / invalid params / negative HTTP / 400 validation:** do **not** add those `test_*` functions to `tests/e2e/online_serving/test_*.py` or `*_expansion.py`. **Move or author them** under `tests/dfx/reliability/invalid_param_test/` in the **endpoint-matching script** (see **Invalid parameter validation** below). Success-path e2e and invalid-param dfx tests must stay in separate modules.
@@ -270,7 +270,7 @@ def test_text_to_video_001(omni_server, openai_client) -> None:

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

-*Performance tests:* add or extend entries under `tests/perf/` (and JSON configs where the project uses them), with **explicit baselines** and the same marker/run-level pairing as the CI step that will execute them.
+*Performance tests:* add or extend entries under `tests/dfx/perf/tests/` (and JSON configs where the project uses them), with **explicit baselines**, **`mark`** on each case (`hardware_marks` + `full_model` + type marker), and the same nightly Perf step pattern as in-tree configs.

 **3. Cross-cutting rules**

@@ -592,27 +592,31 @@ Implementation strategy:
 **When the user explicitly asks for L4 perf (or full L4 including perf):**

 1. **Do not** put throughput/latency assertions inside `test_*_expansion.py` — perf uses the **dfx benchmark harness**, not `omni_server` e2e fixtures.
-2. Add or extend a **JSON config** under `tests/dfx/perf/tests/` (mirror in-tree names: `test_qwen_image_vllm_omni.json`, `test_bagel_vllm_omni.json`).
-3. Each JSON entry: `test_name`, `server_params` (`model`, `serve_args`), `benchmark_params[]` with explicit **`baseline`** (`throughput_qps`, `latency_mean`, `peak_memory_mb_mean`, …).
+2. Add or extend a **JSON config** under `tests/dfx/perf/tests/` (mirror in-tree names: `test_qwen_image_vllm_omni.json`, `test_cosmos3_vllm_omni.json`, `test_qwen3_omni_async_chunk.json`).
+3. Each JSON **case**: `test_name`, optional **`mark`** (`hardware_marks` required when present; `marks`: `full_model` + `omni`/`tts`/`diffusion`), `server_params` (`model`, `serve_args` or `stage_config_name`), `benchmark_params[]` with **`name`** and explicit **`baseline`** (`throughput_qps`, `latency_mean`, `peak_memory_mb_mean`, …). Diffusion cases also set `server_type` (e.g. `"vllm-omni"`) and usually `benchmark_endpoint`.
 4. Run locally via the matching script (model type → runner):

 | Model type | Runner script | Example config |
 |------------|---------------|----------------|
 | **Diffusion X2I/X2V** | `tests/dfx/perf/scripts/run_diffusion_benchmark.py` | `tests/dfx/perf/tests/test_qwen_image_vllm_omni.json` |
 | **TTS** | `tests/dfx/perf/scripts/run_benchmark.py` | `tests/dfx/perf/tests/test_tts.json` (shared) or `test_{slug}.json` (dedicated; use when the model must not join the shared nightly matrix before integration — e.g. `test_voxcpm2.json`, `test_coqui_tts.json`) |
-| **Omni** | (see existing Omni Perf steps — `run_benchmark.py` / dedicated omni configs in-tree) | `tests/dfx/perf/tests/test_omni*.json` |
+| **Omni** | `tests/dfx/perf/scripts/run_benchmark.py` | `test_qwen3_omni_no_async_chunk.json`, `test_qwen3_omni_async_chunk.json`, `test_qwen3_omni_vllm_text.json`, `test_qwen3_omni_multi_replicas.json` |

-5. Wire **`test-nightly.yml` · Perf Test** step for that model (separate from Function Test): export `DIFFUSION_BENCHMARK_DIR` / `BENCHMARK_DIR`, run pytest on the script with `--test-config-file`, **upload artifacts** (`buildkite-agent artifact upload`), often **multi-GPU H100** for diffusion perf.
+5. Wire **`test-nightly.yml` · Perf Test** step for that model (separate from Function Test): export `DIFFUSION_BENCHMARK_DIR` / `BENCHMARK_DIR`, run pytest on the script with **`--test-config-file`** (nightly perf steps do not use `-m`), **upload artifacts** (`buildkite-agent artifact upload`), often **multi-GPU H100** for diffusion perf.
 6. **One benchmark scenario → one `test_name` block** in JSON (same spirit as one `test_*` per function case). Combine server `serve_args` + workload in one entry; use multiple `benchmark_params` rows for size/step sweeps under the same server config.

-**Perf local command (diffusion example):**
+**Perf local commands:**

 ```bash
 cd tests
 export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
 export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+# Single file (CI-like)
 pytest -s -v dfx/perf/scripts/run_diffusion_benchmark.py \
  --test-config-file dfx/perf/tests/test_<model>_vllm_omni.json
+# Bulk load + filter by JSON mark
+pytest -sv dfx/perf/scripts/run_diffusion_benchmark.py -m "full_model and diffusion and H100"
+pytest -sv dfx/perf/scripts/run_benchmark.py -m "full_model and omni and H100"
 ```

 **Clarify in the test plan** which L4 pillars you deliver: `Function only` | `Function + Perf` | `Function + Accuracy + Perf`.
@@ -127,8 +127,12 @@ pytest -s -v e2e/online_serving/test_qwen_image_expansion.py -m "full_model and
 cd tests
 export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
 export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+# CI-like: single JSON file (nightly Perf Test steps)
 pytest -s -v dfx/perf/scripts/run_diffusion_benchmark.py \
  --test-config-file dfx/perf/tests/test_qwen_image_vllm_omni.json
+# Local bulk load: all *.json under tests/dfx/perf/tests/, filter by JSON mark
+pytest -sv dfx/perf/scripts/run_diffusion_benchmark.py -m "full_model and diffusion and H100"
+pytest -sv dfx/perf/scripts/run_benchmark.py -m "full_model and omni and H100"
 ```

 Broad marker sweep (when no explicit file shard):
@@ -150,7 +154,26 @@ Nightly **L4** for a model is often **multiple jobs** in `test-nightly.yml`, not

 **Rule:** “L4 functional cases” → **Function only** by default. State that Perf/Accuracy are separate pillars; add them only when requested.

-**Diffusion perf JSON** (`tests/dfx/perf/tests/test_<slug>_vllm_omni.json`): array of objects with `test_name`, `server_params` (`model`, `serve_args`), `benchmark_params[]`, and per-workload **`baseline`** (`throughput_qps`, `latency_mean`, `peak_memory_mb_mean`). Copy structure from `test_qwen_image_vllm_omni.json` in-tree.
+**Diffusion perf JSON** (`tests/dfx/perf/tests/test_<slug>_vllm_omni.json`): array of **case objects**, each with:
+
+| Field | Required | Notes |
+|-------|----------|-------|
+| `test_name` | Yes | Unique server/workload id |
+| `mark` | Recommended | When present: **`hardware_marks` required** + optional `marks` (`full_model`, `diffusion`, `local_model`, …). Parsed by `resolve_pytest_marks`; enables local `-m` filtering. |
+| `server_type` | Diffusion | e.g. `"vllm-omni"` — routes case to `run_diffusion_benchmark.py` |
+| `benchmark_endpoint` | Diffusion | e.g. `/v1/videos`, `/v1/images/generations` |
+| `server_params` | Yes | `model`, `serve_args`, … |
+| `benchmark_params[]` | Yes | Each row: **`name`** (pytest id suffix), workload fields, **`baseline`** |
+
+Copy structure from `test_qwen_image_vllm_omni.json` or `test_cosmos3_vllm_omni.json` (2-GPU mark example).
+
+**Omni perf JSON**: `test_qwen3_omni_*.json` — same `mark` shape with `"marks": ["full_model", "omni"]`; run via `run_benchmark.py`.
+
+**Runner split**: `is_diffusion_perf_config()` → diffusion when `server_type` is set or `"diffusion"` ∈ `mark.marks`; otherwise omni/tts → `run_benchmark.py`.
+
+**Param ids**: `{test_name}-{benchmark_params.name}` (e.g. `test_tts-p0`, not `test_tts-0`).
+
+**Result filenames**: runtime hardware from `get_runtime_resource_label()`; `H100` omitted on default CI pool.

 **Diffusion perf nightly step** (under **Diffusion X2I(&A&T)**; copy full `kubernetes` plugins from `· Perf Test · Qwen-Image`):

@@ -177,9 +200,9 @@ Nightly **L4** for a model is often **multiple jobs** in `test-nightly.yml`, not

 | Model type | Perf runner | Config example |
 |------------|-------------|----------------|
-| Diffusion X2I/X2V | `dfx/perf/scripts/run_diffusion_benchmark.py` | `test_qwen_image_vllm_omni.json` |
-| TTS | `dfx/perf/scripts/run_benchmark.py` | `test_tts.json` (`BENCHMARK_DIR`) |
-| Omni | `run_benchmark.py` / omni configs in-tree | `test_omni*.json` |
+| Diffusion X2I/X2V | `dfx/perf/scripts/run_diffusion_benchmark.py` | `test_qwen_image_vllm_omni.json`, `test_cosmos3_vllm_omni.json` |
+| TTS | `dfx/perf/scripts/run_benchmark.py` | `test_tts.json`, `test_voxcpm2.json`, `test_higgs_audio_v3.json` |
+| Omni | `dfx/perf/scripts/run_benchmark.py` | `test_qwen3_omni_async_chunk.json`, `test_qwen3_omni_no_async_chunk.json`, … |

 Do **not** put throughput/latency baselines inside `test_*_expansion.py` — that belongs in the dfx perf JSON + nightly Perf job.

@@ -248,6 +248,7 @@ datasets/

 # Output files
 outputs/
+!vllm_omni/outputs/
 results/
 generated/
 output_*/
@@ -16,6 +16,7 @@ Easy, fast, and cheap omni-modality model serving for everyone
 ---

 *Latest News* 🔥
+- [2026/07] We released [0.24.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.24.0) - aligned with the vLLM 0.24 release line, expanding production-ready coverage across TTS, speech, diffusion, image/video generation, and robot-policy serving, with major Omni stage runtime refactoring, diffusion request-level batching, async output materialization, quantization/cache/memory improvements, and broad CUDA/ROCm/XPU/NPU support.
 - [2026/06] We released [0.22.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.22.0) - an **omnimodal world-model** release aligned with vLLM 0.22, featuring [Nvidia Cosmos3](recipes/cosmos3/Cosmos3-Nano.md)/DreamZero world model support, expanded quantization coverage across Blackwell/NPU/XPU, TTS production improvements, new models including MiniCPM-o 4.5, MOSS-TTS, and Lance, plus RL integration with [VeRL-Omni](https://github.com/verl-project/verl-omni).
 - [2026/05] We released [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0) - refreshes the serving/runtime stack for large-scale omni workloads, and improves diffusion model performance, quantization, and hardware readiness across CUDA, ROCm, MUSA, NPU, and XPU backends.
 - [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments.
@@ -31,9 +32,9 @@ Easy, fast, and cheap omni-modality model serving for everyone

 [vLLM](https://github.com/vllm-project/vllm) was originally designed to support large language models for text-based autoregressive generation tasks. vLLM-Omni is a framework that extends its support for omni-modality model inference and serving:

- **Omni-modality**: Text, image, video, and audio data processing
+- **Omni-modality**: Text, image, audio, video, and action data processing
 - **Non-autoregressive Architectures**: extend the AR support of vLLM to Diffusion Transformers (DiT) and other parallel generation models
- **Heterogeneous outputs**: from traditional text generation to multimodal outputs
+- **Heterogeneous outputs**: from traditional text generation to multimodal and action outputs

 <p align="center">
  <picture>
@@ -57,9 +58,10 @@ vLLM-Omni is flexible and easy to use with:

 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:

- **Omni-modality models** (e.g. Qwen3-Omni, Cosmos, HunyuanImage, BAGEL)
+- **Omni-modality models** (e.g. Qwen3-Omni, Cosmos3, HunyuanImage, BAGEL)
 - **TTS models** (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
 - **Diffusion models** — image, video, and audio generation (e.g. Qwen-Image, Wan2.2, FLUX)
+- **Robot-policy and action models** (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)

 ## Getting Started

@@ -28,6 +28,7 @@ class RequestFuncInput:
    slo_ms: float | None = None
    extra_body: dict[str, Any] = field(default_factory=dict)
    image_paths: list[str] | None = None
+    video_paths: list[str] | None = None
    request_id: str = field(default_factory=lambda: str(uuid.uuid4()))
    default_bot_task: str | None = DEFAULT_EDITS_BOT_TASK

@@ -338,6 +339,11 @@ async def async_request_v1_videos(
        form.add_field(k, str(v))

    image_file = None
+    if input.image_paths and input.video_paths:
+        output.error = "Only one of image_paths or video_paths can be provided"
+        output.success = False
+        return output
+
    if input.image_paths and len(input.image_paths) > 0:
        image_path = input.image_paths[0]
        image_file = open(image_path, "rb")
@@ -345,7 +351,16 @@ async def async_request_v1_videos(
            "input_reference",
            image_file,
            filename=os.path.basename(image_path),
-            content_type="application/octet-stream",
+            content_type=_guess_mime_type(image_path),
+        )
+    elif input.video_paths and len(input.video_paths) > 0:
+        video_path = input.video_paths[0]
+        image_file = open(video_path, "rb")
+        form.add_field(
+            "input_reference",
+            image_file,
+            filename=os.path.basename(video_path),
+            content_type=_guess_mime_type(video_path),
        )

    job_id = None
@@ -73,6 +73,10 @@ Usage:
            {"width":854,"height":480,"num_inference_steps":18,"num_frames":120,"fps":24,"weight":1}
        ]'

+    v2v:
+    python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
+        --endpoint /v1/videos --dataset random --task v2v --num-prompts 1 \
+        --height 720 --width 1280 --fps 24 --num-frames 189 --num-inference-steps 35

 """

@@ -542,11 +546,17 @@ class TraceDataset(BaseDataset):
        if not image_paths:
            single = row.get("image_path")
            image_paths = [single] if single else None
+        video_paths = row.get("video_paths")
+        if not video_paths:
+            single_video = row.get("video_path")
+            video_paths = [single_video] if single_video else None

        if not image_paths and self.args.task in ["i2v", "i2i", "ti2v", "ti2i", "it2i"]:
            raise ValueError(
                f"Task {self.args.task} requires image input, but no image_path or image_paths found in trace row."
            )
+        if not video_paths and self.args.task == "v2v":
+            raise ValueError("Task v2v requires video input, but no video_path or video_paths found in trace row.")

        override_w = self.args.width
        override_h = self.args.height
@@ -570,6 +580,7 @@ class TraceDataset(BaseDataset):
            timestamp=timestamp,
            slo_ms=slo_ms,
            image_paths=image_paths,
+            video_paths=video_paths,
            request_id=str(row.get("request_id")) if row.get("request_id") is not None else str(uuid.uuid4()),
        )

@@ -589,6 +600,8 @@ class CustomDataset(BaseDataset):
    - seed (optional): Random seed
    - image_paths (optional): List of input image paths for i2i/i2v/ti2i tasks
    - image_urls (optional): List of input image URLs (alternative to image_paths)
+    - video_paths (optional): List of input video paths for v2v tasks
+    - video_urls (optional): List of input video URLs (alternative to video_paths)

    Example JSONL for ti2i:
    {"prompt": "Add sunset lighting", "width": 1024, "height": 1024, "image_urls": ["https://example.com/image.jpg"]}
@@ -660,6 +673,42 @@ class CustomDataset(BaseDataset):

        return None

+    def _resolve_video_paths(self, item: dict) -> list[str] | None:
+        if "video_paths" in item and item["video_paths"]:
+            paths = item["video_paths"]
+            if isinstance(paths, str):
+                paths = [paths]
+
+            valid_paths = []
+            for path in paths:
+                if os.path.exists(path):
+                    valid_paths.append(path)
+                else:
+                    raise ValueError(f"Video file not found: {path}")
+
+            return valid_paths if valid_paths else None
+
+        if "video_urls" in item and item["video_urls"]:
+            urls = item["video_urls"]
+            if isinstance(urls, str):
+                urls = [urls]
+
+            downloaded_paths = []
+            for url in urls:
+                try:
+                    response = requests.get(url, timeout=30)
+                    response.raise_for_status()
+                    suffix = os.path.splitext(url)[1] or ".mp4"
+                    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+                        tmp.write(response.content)
+                        downloaded_paths.append(tmp.name)
+                except Exception as e:
+                    raise ValueError(f"Failed to download video from {url}: {e}")
+
+            return downloaded_paths if downloaded_paths else None
+
+        return None
+
    def __len__(self) -> int:
        return len(self.data)

@@ -672,11 +721,22 @@ class CustomDataset(BaseDataset):
        height = item.get("height", self.args.height)
        num_inference_steps = item.get("num_inference_steps", self.args.num_inference_steps)
        seed = item.get("seed", self.args.seed)
-        reserved_keys = {"prompt", "width", "height", "num_inference_steps", "seed", "image_paths", "image_urls"}
+        reserved_keys = {
+            "prompt",
+            "width",
+            "height",
+            "num_inference_steps",
+            "seed",
+            "image_paths",
+            "image_urls",
+            "video_paths",
+            "video_urls",
+        }
        extra_body = {k: v for k, v in item.items() if k not in reserved_keys}

        # Handle image paths/URLs
        image_paths = self._resolve_image_paths(item)
+        video_paths = self._resolve_video_paths(item)

        return RequestFuncInput(
            prompt=prompt,
@@ -684,6 +744,7 @@ class CustomDataset(BaseDataset):
            model=self.model,
            seed=seed,
            image_paths=image_paths,
+            video_paths=video_paths,
            extra_body=extra_body,
            width=width,
            height=height,
@@ -737,6 +798,10 @@ class RandomDataset(BaseDataset):
            self._random_image_path = self._generate_random_image_paths()
        else:
            self._random_image_path = None
+        if self.args.task == "v2v":
+            self._random_video_path = self._generate_random_video_paths()
+        else:
+            self._random_video_path = None

    def __len__(self) -> int:
        return self.num_prompts
@@ -756,13 +821,15 @@ class RandomDataset(BaseDataset):
        if self._sampled_requests:
            profile = self._sampled_requests[idx]
            params.update(profile)
+        prompt = str(params.pop("prompt", f"Random prompt {idx} for benchmarking diffusion models"))
        return RequestFuncInput(
-            prompt=f"Random prompt {idx} for benchmarking diffusion models",
+            prompt=prompt,
            api_url=self.api_url,
            model=self.model,
            seed=self.args.seed,
            extra_body=extra_body,
            image_paths=self._random_image_path,
+            video_paths=self._random_video_path,
            **params,
        )

@@ -781,6 +848,35 @@ class RandomDataset(BaseDataset):
            image_paths.append(image_path)
        return image_paths

+    def _generate_random_video_paths(self) -> list[str]:
+        video_path = os.path.join(tempfile.gettempdir(), "diffusion_benchmark_random_video.mp4")
+        try:
+            import cv2
+
+            profile = self.random_request_config[0] if self.random_request_config else {}
+            width = int(self.args.width or profile.get("width") or 1280)
+            height = int(self.args.height or profile.get("height") or 720)
+            num_frames = int(self.args.num_frames or profile.get("num_frames") or 16)
+            fps = float(self.args.fps or profile.get("fps") or 8)
+            writer = cv2.VideoWriter(video_path, cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height))
+            if not writer.isOpened():
+                raise RuntimeError("cv2.VideoWriter failed to open")
+            for frame_idx in range(num_frames):
+                frame = np.zeros((height, width, 3), dtype=np.uint8)
+                x = int((frame_idx / max(num_frames - 1, 1)) * max(width - 96, 1))
+                y = height // 2
+                frame[:, :, 0] = 24
+                frame[:, :, 1] = 32
+                frame[:, :, 2] = 48
+                frame[max(y - 36, 0) : min(y + 36, height), x : min(x + 96, width), :] = (64, 128, 220)
+                writer.write(frame)
+            writer.release()
+            return [video_path]
+        except Exception as e:
+            raise RuntimeError(
+                "Failed to generate synthetic v2v input video. Install opencv-python or provide a custom dataset."
+            ) from e
+

 def _compute_expected_latency_ms_from_base(req: RequestFuncInput, args, base_time_ms: float | None) -> float | None:
    """Compute expected execution time (ms) based on a base per-step-per-frame unit time.
@@ -1112,7 +1208,7 @@ def wait_for_service(base_url: str, timeout: int = 120) -> None:


 def _default_endpoint_for_task(task: str) -> str:
-    if task in {"t2v", "i2v", "ti2v"}:
+    if task in {"t2v", "i2v", "ti2v", "v2v"}:
        return "/v1/videos"
    if task in {"i2i", "ti2i", "it2i"}:
        return "/v1/images/edits"
@@ -1126,7 +1222,7 @@ async def benchmark(args):
    if args.base_url is None:
        args.base_url = f"http://{args.host}:{args.port}"

-    VIDEO_TASKS = {"t2v", "i2v", "ti2v"}
+    VIDEO_TASKS = {"t2v", "i2v", "ti2v", "v2v"}
    IMAGE_TASKS = {"t2i", "i2i", "ti2i", "it2i"}

    if args.task in VIDEO_TASKS:
@@ -1177,6 +1273,10 @@ async def benchmark(args):
    if args.return_stage_metrics and args.endpoint in _STAGE_METRICS_ENDPOINTS:
        for req in requests_list:
            req.extra_body.setdefault(_RETURN_STAGE_METRICS_FIELD, True)
+    if args.extra_body:
+        extra_body = json.loads(args.extra_body)
+        for req in requests_list:
+            req.extra_body.update(extra_body)

    if args.endpoint == "/v1/images/edits":
        for req in requests_list:
@@ -1328,7 +1428,7 @@ if __name__ == "__main__":
        "--task",
        type=str,
        default="t2v",
-        choices=["t2v", "i2v", "ti2v", "ti2i", "i2i", "it2i", "t2i"],
+        choices=["t2v", "i2v", "ti2v", "v2v", "ti2i", "i2i", "it2i", "t2i"],
        help="Task type.",
    )
    parser.add_argument(
@@ -1443,6 +1543,12 @@ if __name__ == "__main__":
            '{"width":768,"height":768,"num_inference_steps":20,"weight":0.85}]'
        ),
    )
+    parser.add_argument(
+        "--extra-body",
+        type=str,
+        default=None,
+        help="JSON object merged into every generated request body/form.",
+    )
    parser.add_argument(
        "--num-input-images",
        type=int,
@@ -2,7 +2,7 @@
 # The pinned VLLM_PRECOMPILED_WHEEL_COMMIT and VLLM_BASE_TAG below must
 # be updated on every rebase to match the new upstream vllm commit/tag.
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.24.0
+ARG VLLM_BASE_TAG=v0.25.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}
@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai:v0.24.0
+ARG BASE_IMAGE=vllm/vllm-openai:v0.25.0
 FROM ${BASE_IMAGE}

 ARG COMMON_WORKDIR=/app
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.24.0
+ARG VLLM_ASCEND_TAG=v0.25.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 # WORKDIR /vllm-workspace/vllm
@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.24.0-a3
+ARG VLLM_ASCEND_TAG=v0.25.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 # WORKDIR /vllm-workspace/vllm
@@ -1,11 +1,13 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.24.0
+ARG VLLM_ASCEND_TAG=v0.25.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 ARG APP_DIR=/vllm-workspace/vllm-omni
 WORKDIR ${APP_DIR}
 COPY . .

+RUN sed -i "s@^\(deb.*\)https\?://[a-z0-9.-]*\.ubuntu\.com@\1http://cache-service.nginx-pypi-cache.svc.cluster.local:8081@g" /etc/apt/sources.list
+
 RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        ffmpeg \
@@ -1,11 +1,13 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.24.0-a3
+ARG VLLM_ASCEND_TAG=v0.25.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}

 ARG APP_DIR=/vllm-workspace/vllm-omni
 WORKDIR ${APP_DIR}
 COPY . .

+RUN sed -i "s@^\(deb.*\)https\?://[a-z0-9.-]*\.ubuntu\.com@\1http://cache-service.nginx-pypi-cache.svc.cluster.local:8081@g" /etc/apt/sources.list
+
 RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        ffmpeg \
@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.24.0
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.25.0
 FROM ${BASE_IMAGE} AS base

 # Declare a variable to know if we want to use the nightly build or the stable build.
@@ -88,7 +88,7 @@ ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE="copy"

-ARG VLLM_VERSION=v0.24.0
+ARG VLLM_VERSION=v0.25.0
 RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
 WORKDIR /workspace/vllm

@@ -23,7 +23,7 @@ nav:
      - Image-To-Video: user_guide/examples/offline_inference/image_to_video.md
      - InternVLA-A1: user_guide/examples/offline_inference/internvla_a1.md
      - MammothModa2-Preview: user_guide/examples/offline_inference/mammothmodal2_preview.md
-      - MiMo-Audio Offline Inference: user_guide/examples/offline_inference/mimo_audio.md
+      - MiMo-Audio: user_guide/examples/offline_inference/mimo_audio.md
      - Qwen2.5-Omni: user_guide/examples/offline_inference/qwen2_5_omni.md
      - Qwen3-Omni: user_guide/examples/offline_inference/qwen3_omni.md
      - Text-To-Audio: user_guide/examples/offline_inference/text_to_audio.md
@@ -38,7 +38,7 @@ nav:
      - GLM-Image Online Serving: user_guide/examples/online_serving/glm_image.md
      - Image-To-Image: user_guide/examples/online_serving/image_to_image.md
      - Image-To-Video: user_guide/examples/online_serving/image_to_video.md
-      - Online serving Example of vLLM-Omni for MiMo-Audio: user_guide/examples/online_serving/mimo_audio.md
+      - MiMo-Audio: user_guide/examples/online_serving/mimo_audio.md
      - Qwen2.5-Omni: user_guide/examples/online_serving/qwen2_5_omni.md
      - Qwen3-Omni: user_guide/examples/online_serving/qwen3_omni.md
      - Text-To-Speech: user_guide/examples/online_serving/text_to_speech.md
@@ -80,6 +80,7 @@ nav:
      - Quantized KV Cache: user_guide/quantization/quantized_kvcache.md
      - FP8 W8A8: user_guide/quantization/fp8.md
      - Int8 W8A8: user_guide/quantization/int8.md
+      - BitsAndBytes W4: user_guide/quantization/bitsandbytes.md
      - MXFP8 W8A8: user_guide/quantization/mxfp8.md
      - MXFP4 W4A4: user_guide/quantization/mxfp4.md
      - ModelOpt: user_guide/quantization/modelopt.md
@@ -28,9 +28,9 @@ Easy, fast, and cheap omni-modality model serving for everyone

 [vLLM](https://github.com/vllm-project/vllm) was originally designed to support large language models for text-based autoregressive generation tasks. vLLM-Omni is a framework that extends its support for omni-modality model inference and serving:

- **Omni-modality**: Text, image, video, and audio data processing
+- **Omni-modality**: Text, image, audio, video, and action data processing
 - **Non-autoregressive Architectures**: extend the AR support of vLLM to Diffusion Transformers (DiT) and other parallel generation models
- **Heterogeneous outputs**: from traditional text generation to multimodal outputs
+- **Heterogeneous outputs**: from traditional text generation to multimodal and action outputs

 <p align="center">
  <picture>
@@ -55,8 +55,10 @@ vLLM-Omni is flexible and easy to use with:

 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:

- Omni-modality models (e.g. Qwen2.5-Omni, Qwen3-Omni)
- Multi-modality generation models (e.g. Qwen-Image)
+- Omni-modality models (e.g. Qwen3-Omni, Cosmos3, HunyuanImage, BAGEL)
+- TTS models (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
+- Diffusion models — image, video, and audio generation (e.g. Qwen-Image, Wan2.2, FLUX)
+- Robot-policy and action models (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)

 For more information, checkout the following:

@@ -46,17 +46,17 @@ Engine classes for offline and online inference.
 - [vllm_omni.engine.PromptEmbedsPayload][]
 - [vllm_omni.engine.arg_utils.OmniEngineArgs][]
 - [vllm_omni.engine.async_omni_engine.AsyncOmniEngine][]
- [vllm_omni.engine.mm_outputs.MultimodalCompletionOutput][]
- [vllm_omni.engine.mm_outputs.MultimodalPayload][]
 - [vllm_omni.engine.orchestrator.Orchestrator][]
 - [vllm_omni.engine.orchestrator.OrchestratorRequestState][]
- [vllm_omni.engine.output_modality.OutputModality][]
- [vllm_omni.engine.output_modality.TensorAccumulationStrategy][]
- [vllm_omni.engine.output_processor.MultimodalOutputProcessor][]
- [vllm_omni.engine.output_processor.OmniRequestState][]
 - [vllm_omni.engine.stage_engine_core_client.StageEngineCoreClient][]
 - [vllm_omni.engine.stage_init_utils.StageMetadata][]
 - [vllm_omni.engine.stage_init_utils.StartedLlmStage][]
+- [vllm_omni.outputs.mm_outputs.MultimodalCompletionOutput][]
+- [vllm_omni.outputs.mm_outputs.MultimodalPayload][]
+- [vllm_omni.outputs.output_modality.OutputModality][]
+- [vllm_omni.outputs.output_modality.TensorAccumulationStrategy][]
+- [vllm_omni.outputs.output_processor.MultimodalOutputProcessor][]
+- [vllm_omni.outputs.output_processor.OmniRequestState][]

 ## Core

@@ -18,11 +18,17 @@ Specify the port:
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091
 ```

-If you have custom stage configs file, launch the server with command below
+For a migrated model, load a custom deploy configuration with `--deploy-config`:
+
 ```bash
-vllm serve Qwen/Qwen2.5-Omni-7B --omni --stage-configs-path /path/to/stage_configs_file
+vllm serve Qwen/Qwen2.5-Omni-7B --omni --deploy-config /path/to/deploy_config.yaml
 ```

+The deprecated `--stage-configs-path` flag is retained for models that still use the legacy `stage_args` schema:
+
+```bash
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --stage-configs-path /path/to/legacy_stage_config.yaml
+```

 ## bench

@@ -2,7 +2,7 @@
 The vllm bench command launches the vLLM-Omni benchmark to evaluate the performance of multimodal models.

 ## Notes
-We currently only support using the "openai-chat-omni" and "openai-image-edits-omni" backend.
+vLLM-Omni registers the `openai-chat-omni`, `openai-audio-speech`, `openai-image-edits-omni`, and `daily-omni` serving benchmark backends.

 ## Basic Parameter Description
 You can use `vllm bench serve --omni --help=all` to get descriptions of all parameters. The commonly used parameters are described below:
@@ -10,7 +10,7 @@ You can use `vllm bench serve --omni --help=all` to get descriptions of all para
  Enable Omni (multimodal) mode, supporting multimodal inputs and outputs such as images, videos, and audio.

 - `--backend`
-  Specify the backend adapter as openai-chat-omni, using OpenAI Chat compatible API behavior as the protocol. Currently only openai-chat-omni is supported.
+  Specify the backend adapter. vLLM-Omni adds `openai-chat-omni`, `openai-audio-speech`, `openai-image-edits-omni`, and `daily-omni` to the upstream vLLM backend choices.

 - `--model`
  The model identifier to load, filled according to the models supported by vLLM-Omni.
@@ -55,7 +55,7 @@ You can use `vllm bench serve --omni --help=all` to get descriptions of all para
                    'Allowed metric names are "ttft", "tpot", "itl", "ttfc", "tpoc", "icl", '
                    '"tpop", "e2el", "audio_ttfp", "audio_rtf", "audio_duration". '

- `-print-stage`
+- `--print-stage`
 Print per-stage benchmark metrics for --omni serving when stage metrics are returned by the server. Disabled by default.

 - `--save-result`
@@ -312,7 +312,7 @@ vllm bench serve \
  --ignore-eos \
  --percentile-metrics ttft,tpot,itl \
  --random-output-len 2 \
-  --extra_body '{"modalities": ["text"]}'
+  --extra-body '{"modalities": ["text"]}'
 ```

 If successful, you will see the following output:
@@ -390,7 +390,7 @@ vllm bench serve --omni \
  --ignore-eos \
  --percentile-metrics ttft,tpot,itl,e2el,audio_ttfp,audio_rtf,ttfc,tpoc,icl \
  --random-output-len 900 \
-  --extra_body '{"modalities": ["text","audio"]}' \
+  --extra-body '{"modalities": ["text","audio"]}' \
  --print-stage
 ```

@@ -92,12 +92,12 @@ Full pattern hierarchy: `TakeRank`, `Union`, `GatherDim`, `AllGather`, `StitchSp

 From highest to lowest, when the same axis is touched by multiple layers:

-1. **CLI overrides** (`--stage_N_tensor_parallel_size`, `--stage-overrides`, `--omni-lb-policy`, etc.). A CLI value always wins; a warning is emitted whenever it overrides a strategy-declared axis (see `_reconcile_strategy_with_cli` in `vllm_omni/config/config_factory.py`).
+1. **CLI overrides** (`--tensor-parallel-size`, `--stage-overrides`, `--omni-lb-policy`, etc.). A CLI value always wins; a warning is emitted whenever it overrides a strategy-declared axis (see `_reconcile_strategy_with_cli` in `vllm_omni/config/config_factory.py`).
 2. **Strategy YAML** (`--strategy-config`). Derived sizing is written onto the merged stages with *conflict-on-explicit* semantics — if the deploy YAML already set a value to something different, application raises `StrategyApplyError`.
 3. **Deploy YAML** (`--deploy-config` or the bundled `vllm_omni/deploy/<model_type>.yaml`).
 4. **Parser defaults.**

-After CLI overrides land, the device-layout guard re-runs (`check_device_layout`, called from `_reconcile_strategy_with_cli`) against the *effective* `tp * dp * pp * num_replicas` and the resolved `devices` string, so a `--devices` or `--tensor-parallel-size` that conflicts with the strategy's world size cannot slip past the pre-spawn check.
+After CLI overrides land, the device-layout guard re-runs (`check_device_layout`, called from `_reconcile_strategy_with_cli`) against the *effective* `tp * dp * pp * num_replicas` and the resolved `devices` string, so a `devices` value passed through `--stage-overrides` or a `--tensor-parallel-size` value that conflicts with the strategy's world size cannot slip past the pre-spawn check.

 ## Worked example

@@ -214,7 +214,7 @@ Strategy keys MUST be `model_stage` *names* (strings), the same labels stage con
    Use the `model_stage` name (e.g. `thinker`) instead.

 !!! warning "Mixing CLI overrides with a strategy on the same axis"
-    A strategy is meant to be the single writer for the axes it declares. If a CLI override (`--stage_0_tensor_parallel_size`, `--stage-overrides '{"0": {...}}'`) sets the same axis, the **CLI value wins** and a warning is emitted from `_reconcile_strategy_with_cli`. The device-layout guard then re-runs against the effective layout, so a CLI override that breaks `tp * dp * pp * num_replicas` will still fail fast at config time.
+    A strategy is meant to be the single writer for the axes it declares. If a CLI override (`--tensor-parallel-size 2`, `--stage-overrides '{"0": {"tensor_parallel_size": 2}}'`) sets the same axis, the **CLI value wins** and a warning is emitted from `_reconcile_strategy_with_cli`. The device-layout guard then re-runs against the effective layout, so a CLI override that breaks `tp * dp * pp * num_replicas` will still fail fast at config time.

 !!! warning "Passing `--strategy-config` against the legacy `--stage-configs-path`"
    Composable parallel only applies on the registry-based deploy path. If a model still resolves through the legacy `stage_args` YAML (loaded via `--stage-configs-path`), `--strategy-config` is silently inapplicable and emits:
@@ -1,11 +1,11 @@
-# Stage configs for vLLM-Omni
+# Deploy and legacy stage configurations

-In vLLM-Omni, the target model is separated into multiple stages, which are processed by different LLMEngines, DiffusionEngines or other types of engines. Depending on different types of stages, such as Autoregressive (AR) stage or Diffusion transformer (DiT) stage, each can choose corresponding schedulers, model workers to load with the Engines in a plug-in fashion.
+In vLLM-Omni, a model's `PipelineConfig` defines its fixed stage topology, while a deploy configuration controls how those stages run. Models that have not migrated to this split still use legacy stage configurations with a `stage_args` schema.

 !!! note
    Default deploy config YAMLs (for example, `vllm_omni/deploy/qwen2_5_omni.yaml`, `vllm_omni/deploy/qwen3_omni_moe.yaml`, and `vllm_omni/deploy/qwen3_tts.yaml`) are bundled and loaded automatically when neither `--stage-configs-path` nor `--deploy-config` is provided — the model registry resolves the right pipeline + deploy YAML by `model_type`. The bundled defaults have been verified on 1xH100 for Qwen2.5-Omni and 2xH100 for Qwen3-Omni. Models that have not yet migrated to the new schema continue to use the legacy `vllm_omni/model_executor/stage_configs/<model>.yaml` files via `--stage-configs-path`.

-## New deploy schema reference
+## Deploy configuration schema

 The new deploy schema lives under `vllm_omni/deploy/` and is paired with a frozen `PipelineConfig` registered by the model's `pipeline.py`. Each deploy YAML has these top-level fields:

@@ -82,14 +82,14 @@ stages:
    input_connectors:  {from_stage_0: shm}
 ```

-### CLI flags introduced in this refactor
+### CLI flags

 | Flag | Description |
 |------|-------------|
-| `--deploy-config PATH` | Load a new-schema deploy YAML. Takes precedence over `--stage-configs-path`. **Optional** — when omitted, the bundled `vllm_omni/deploy/<model_type>.yaml` is auto-loaded by the model registry. |
+| `--deploy-config PATH` | Load a new-schema deploy YAML. It is mutually exclusive with `--stage-configs-path`. **Optional** — when omitted, the bundled `vllm_omni/deploy/<model_type>.yaml` is auto-loaded by the model registry. |
 | `--stage-overrides JSON` | Per-stage JSON overrides, e.g. `'{"0":{"gpu_memory_utilization":0.5}}'`. Per-stage values always win over global flags. |
 | `--async-chunk` / `--no-async-chunk` | Flip the deploy YAML's `async_chunk:` bool. Unset (default) leaves the YAML value in force. |
-| `--stage-configs-path` | **Deprecated.** Accepts legacy `stage_args` yamls and (auto-detected) new deploy yamls; emits a deprecation warning. Migrate to `--deploy-config`. To be removed in a follow-up PR. |
+| `--stage-configs-path` | **Deprecated.** Load a legacy `stage_args` YAML. New-schema YAMLs passed to this flag are auto-detected for compatibility, but should use `--deploy-config`. It is mutually exclusive with `--deploy-config` and will be removed in a future release. |

 ### Stage-Based CLI Paradigm

@@ -146,7 +146,7 @@ CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni \

 From highest to lowest:

-1. Per-stage flags (`--stage-overrides` JSON, `--stage-<id>-<key>` if registered)
+1. Per-stage overrides (`--stage-overrides` JSON)
 2. Explicit global CLI flags (`--gpu-memory-utilization 0.85`, etc.)
 3. Platform section (`platforms.npu.stages`, etc.) on top of the base `stages:`
 4. Overlay YAML (via `base_config:`) on top of the base YAML
@@ -210,7 +210,7 @@ Effective config per stage after the merge:
 | 1 | `max_model_len` | `16384` | global CLI |
 | 2 | (all defaults) | — | base YAML (no overrides apply) |

-Therefore, as a core part of vLLM-Omni, the stage configs for a model have several main functions:
+Therefore, as a core part of vLLM-Omni, a model's pipeline and deployment configurations have several main functions:

 - Claim partition of stages and their corresponding class implementation in `model_executor/models`.
 - The disaggregated configuration for each stage and the communication topology among them.
@@ -219,16 +219,16 @@ Therefore, as a core part of vLLM-Omni, the stage configs for a model have sever
 - Default input parameters.

 To override specific parameters, explicitly inject the customized configuration schema
-in both online and offline instantiation flows. Prioritize the `--deploy-config` flag
-when loading the new-schema deploy YAML schemas, reserving the `--stage-configs-path` parameter
+in both online and offline instantiation flows. Use the `--deploy-config` flag
+when loading a deploy configuration, reserving the `--stage-configs-path` parameter
 exclusively to maintain compatibility with legacy `stage_args` YAML constructs.

 Examples:

-For offline (Assume necessary dependencies have been imported):
+For offline inference (assuming the necessary dependencies have been imported):
 ```python
 model_name = "Qwen/Qwen2.5-Omni-7B"
-omni = Omni(model=model_name, stage_configs_path="/path/to/custom_stage_configs.yaml")
+omni = Omni(model=model_name, deploy_config="/path/to/deploy_config.yaml")
 ```

 For online serving:
@@ -242,11 +242,14 @@ Legacy online serving:
 vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
 ```
 !!! important
-    We are actively iterating on the definition of stage configs, and we welcome all feedbacks from both community users and developers to help us shape the development!
+    We are actively iterating on the definition of deployment configurations, and we welcome feedback from users and developers.
+
+## Legacy stage configuration reference
+
+The remainder of this page documents the deprecated `stage_args` schema for models that have not migrated to deploy configurations. The following historical Qwen2.5-Omni configuration illustrates that schema; current Qwen2.5-Omni deployments should use `vllm_omni/deploy/qwen2_5_omni.yaml` instead.

-Below is a specific example of stage_configs.yaml in Qwen2.5-omni.
 ```python
-# stage config for running qwen2.5-omni with AsyncOmniEngine + Orchestrator runtime.
+# Historical stage configuration for Qwen2.5-Omni.
 stage_args:
  - stage_id: 0 # mark the unique id for each stage
    runtime: # The disaggregated configuration
@@ -338,7 +341,7 @@ runtime:

 ```

-## Stage Configuration Arguments
+## Legacy stage configuration arguments

 Each stage in the `stage_args` list contains the following configuration options:

@@ -67,6 +67,21 @@ docs/
 └── stylesheets/         # Custom CSS
 ```

+## Naming Model Examples
+
+Offline inference and online serving examples for the same model use the same
+directory name and the shared display name in
+`examples/model_display_names.yml`. Use these title forms:
+
+- `# <Model>: Offline inference`
+- `# <Model>: Online serving`
+
+The documentation generator uses the full title for the page H1 and the shared
+display name alone for navigation, and fails the build if a mapped README uses
+a different H1. Keep checkpoint identifiers in commands and prose rather than
+in the display name, and do not rename an existing example directory solely to
+adjust its title because the directory defines its public documentation URL.
+
 ## Publishing Documentation

 ### GitHub Pages (Recommended)
@@ -126,11 +126,12 @@ Through five levels (L1-L5) and common (Common) specifications, the system clari
        <strong>Interface tests:</strong><br>
        /tests/entrypoints/test_xxx<br>
        <strong>Performance:</strong><br>
-        /tests/dfx/perf/tests/test_qwen_omni.json (Omni), test_tts.json (TTS),<br>
-        and /tests/dfx/perf/tests/test_{diffusion_model}_vllm_omni.json (Diffusion)<br>
+        /tests/dfx/perf/tests/test_qwen3_omni_*.json (Omni), test_tts.json (TTS),<br>
+        test_voxcpm2.json, test_higgs_audio_v3.json, and<br>
+        /tests/dfx/perf/tests/test_{diffusion_model}_vllm_omni.json (Diffusion)<br>
        <strong>Doc Test:</strong><br>
-        tests/example/online_serving/test_{model_name}.py<br>
-        tests/example/offline_inference/test_{model_name}.py<br>
+        tests/examples/online_serving/test_{model_name}.py<br>
+        tests/examples/offline_inference/test_{model_name}.py<br>
        <strong>Accuracy Test:</strong><br>
        /tests/e2e/accuracy/test_{model_name}.py
      </td>
@@ -708,16 +709,16 @@ L4 level testing is a comprehensive quality audit before a version release. It e
 ### 3.2 Testing Content and Scope

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

 ### 3.3 Test Directory and Execution Files

 -   ***Functional Testing***: Same directories as L3.
-   ***Performance Test Configuration***: `tests/dfx/perf/tests/test_qwen_omni.json`, `tests/dfx/perf/tests/test_tts.json`, and diffusion configs `tests/dfx/perf/tests/test_*_vllm_omni.json` (e.g. `test_qwen_image_vllm_omni.json`)
+-   ***Performance Test Configuration***: `tests/dfx/perf/tests/test_qwen3_omni_*.json`, `test_tts.json`, `test_voxcpm2.json`, `test_higgs_audio_v3.json`, and diffusion configs `tests/dfx/perf/tests/test_*_vllm_omni.json` (e.g. `test_qwen_image_vllm_omni.json`, `test_cosmos3_vllm_omni.json`). Optional per-case `mark` for local `-m` filtering is documented in Section 3.4 Performance Tests.
 -   ***Documentation Example Tests***:
-   -   `tests/example/online_serving/test_{model_name}.py`
-    -   `tests/example/offline_inference/test_{model_name}.py`
+    -   `tests/examples/online_serving/test_{model_name}.py`
+    -   `tests/examples/offline_inference/test_{model_name}.py`

 ### 3.4 Execution Method and Example

@@ -1,14 +1,41 @@
-When you want to add L4-level ***performance test*** cases, you can refer to the following format for case addition in `tests/dfx/perf/tests/test_qwen_omni.json`, `tests/dfx/perf/tests/test_tts.json`, or diffusion configs such as `tests/dfx/perf/tests/test_*_vllm_omni.json` (selected via `pytest ... run_benchmark.py --test-config-file <path>`):
+When you want to add L4-level ***performance test*** cases, add entries to JSON files under `tests/dfx/perf/tests/` and run them via `tests/dfx/perf/scripts/run_benchmark.py` (omni / TTS) or `run_diffusion_benchmark.py` (diffusion).
+
+**Config file layout (in-tree examples)**
+
+| Model type | Runner | Example JSON files |
+| ---------- | ------ | ------------------ |
+| Omni | `run_benchmark.py` | `test_qwen3_omni_no_async_chunk.json`, `test_qwen3_omni_async_chunk.json`, `test_qwen3_omni_vllm_text.json`, `test_qwen3_omni_multi_replicas.json` |
+| TTS | `run_benchmark.py` | `test_tts.json`, `test_voxcpm2.json`, `test_higgs_audio_v3.json` |
+| Diffusion | `run_diffusion_benchmark.py` | `test_qwen_image_vllm_omni.json`, `test_bagel_vllm_omni.json`, `test_wan22_i2v_vllm_omni.json`, `test_cosmos3_vllm_omni.json`, … |
+
+**How runners pick cases**
+
+Without **`--test-config-file`**, each runner scans all `*.json` under `tests/dfx/perf/tests/` but only keeps its own model type:
+
+- **`run_benchmark.py`**: omni and TTS cases only (skips diffusion JSON).
+- **`run_diffusion_benchmark.py`**: diffusion cases only (skips omni / TTS JSON).
+
+Diffusion cases are detected when the JSON has `server_type` (typically `"vllm-omni"`) or `"diffusion"` in the `mark` array. Omni / TTS JSON has neither.
+
+**Running perf cases**
+
+Pass **`--test-config-file`** to run one JSON file as-is, or omit it for the bulk scan above and filter with pytest **`-m`** on each case's JSON `mark` (for example `pytest … run_benchmark.py -m "full_model and tts and H100"`).

 ```JSON
 {
-    "test_name": "test_qwen3_omni",
+    "test_name": "test_qwen3_omni_async_chunk",
+    "mark": [
+        {"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}},
+        "full_model",
+        "omni"
+    ],
    "server_params": {
        "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
        "stage_config_name": "qwen3_omni.yaml"
    },
    "benchmark_params": [
        {
+            "name": "random_10p",
            "dataset_name": "random",
            "num_prompts": [10, 20],
            "max_concurrency": [1, 4],
@@ -33,8 +60,95 @@ When you want to add L4-level ***performance test*** cases, you can refer to the
 | Field            | Required | Description                                                     |
 | ---------------- | -------- | --------------------------------------------------------------- |
 | test_name        | Yes      | Unique identifier for the test case                             |
+| mark             | No       | Pytest marks for this case (see **`mark` field** below). Omit only for configs not meant to be filtered by `-m`. |
 | server_params    | Yes      | Server-side configuration parameters                            |
 | benchmark_params | Yes      | Benchmark running parameters (supports multiple configurations) |
+| server_type      | Diffusion only | When set (e.g. `"vllm-omni"`), routes the case to `run_diffusion_benchmark.py`. |
+| benchmark_endpoint | Diffusion only | API path for the benchmark client (e.g. `/v1/videos`, `/v1/images/generations`). |
+
+**`mark` field**
+
+Optional top-level field on each perf JSON **case object** (one per `test_name`). `run_benchmark.py` and `run_diffusion_benchmark.py` read it via `tests.dfx.conftest.resolve_pytest_marks` and attach the marks to the corresponding `pytest.param`, so you can filter locally with `-m`.
+
+When `mark` is present, it must be an **array** with exactly one ``hardware_marks`` object (same shape as `@hardware_test` / `hardware_marks()` in `tests/helpers/mark.py`), followed by registered pytest marker name strings such as `full_model`, `omni`, `tts`, `diffusion`, or `local_model`.
+
+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 (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"`. |
+
+Recommended for L4 perf cases:
+
+```JSON
+{
+    "test_name": "test_bagel_single_device_single_stage_t2i",
+    "mark": [
+        {"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 1}},
+        "full_model",
+        "diffusion"
+    ],
+    "server_type": "vllm-omni",
+    "server_params": { "...": "..." },
+    "benchmark_params": [ { "name": "1024x1024_steps20", "...": "..." } ]
+}
+```
+
+Multi-GPU diffusion (example: Cosmos3 with `cfg-parallel-size=2`):
+
+```JSON
+{
+    "test_name": "test_cosmos3_t2i_official_demo_2gpu",
+    "mark": [
+        {"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}},
+        "full_model",
+        "diffusion"
+    ],
+    "server_type": "vllm-omni",
+    "benchmark_endpoint": "/v1/images/generations",
+    "server_params": { "...": "..." },
+    "benchmark_params": [ { "name": "1024x1024_steps4", "...": "..." } ]
+}
+```
+
+- Omni perf: include `"omni"` in the `mark` array
+- TTS perf: include `"tts"` in the `mark` array
+- Diffusion perf: include `"diffusion"` in the `mark` array
+- HunyuanImage local-weight cases: add `"local_model"` to the `mark` array
+
+**Parametrization IDs**
+
+Each `(server, benchmark index)` pair becomes one `pytest.param` with id `{test_name}-{suffix}`. The suffix comes from `benchmark_params[].name` when set (for example `test_tts-p0`, `test_omni-p1`); otherwise it is derived from fields like `task` / `eval_phase`.
+
+**Benchmark result filenames**
+
+Result files use the **runtime** hardware label from `get_runtime_resource_label()` (detected GPU/NPU name on the machine that ran the job), **not** `mark.hardware_marks.res`. On the default H100 CI pool, `H100` is **omitted** from filenames (`resource_label_for_filename`).
+
+Examples:
+
+- Omni/TTS: `result_{test_name}_{optional_hw}_{dataset}_....json` under `BENCHMARK_DIR`
+- Diffusion: one aggregate `diffusion_result_{config_stem}_{optional_hw}_{timestamp}.json` per source JSON file (array of all runs from that file)
+
+**Local commands**
+
+```bash
+# Bulk load + filter by JSON mark
+pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m "full_model and H100 and diffusion"
+pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "full_model and omni and H100"
+
+# Single file (same as nightly CI Perf steps)
+pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py \
+  --test-config-file tests/dfx/perf/tests/test_bagel_vllm_omni.json
+pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
+  --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
+
+# Optional baseline assertion (default off)
+pytest -sv tests/dfx/perf/scripts/run_diffusion_benchmark.py --assert-baseline \
+  --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
+```
+
+See also [Markers for Tests](./tests_markers.md) for registered hardware markers (`H100`, `L4`, `cuda`, `distributed_cuda`, …).

 **`server_params` Configuration**

@@ -79,10 +193,12 @@ You can add any benchmark running parameters you need here. For all optional par
 1.  Change the --xxx-xx-xx running parameters to xxx_xx_xx format and fill them as keys in the JSON file.
 2.  For boolean variables in the running parameters, modify them to forms such as ignore_eos: true/false and fill them into the JSON file.
 3.  Optionally add a `baseline` object (see **Baseline thresholds** below). If you omit `baseline` or leave it empty, the performance test still runs but does not assert metric thresholds from this field.
-4.  The qps and concurrency modes are recommended to be mutually exclusive. For detailed explanations, see the table below:
+4.  Set `"name"` on each `benchmark_params` entry for stable pytest ids and readable result keys.
+5.  The qps and concurrency modes are recommended to be mutually exclusive. For detailed explanations, see the table below:

 | Parameter       | Type        | Required | Example/Values  | Description                                                                                                                                                                                                                                                          |
 | --------------- | ----------- | -------- | --------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| name            | string      | No       | `"1024x1024_steps4"` | Stable suffix for pytest param id and logs. |
 | num_prompts     | int / array | Yes      | 10,[10, 20, 30] | Number of requests. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of qps or max_concurrency, e.g., [10,10,10]. If an array is used, its length must match the number of qps or max_concurrency. |
 | request_rate    | float / array | No  | 0.5, [0.5, 1, inf] | Queries per second. Supports single values or arrays. If a single value is used, it will be automatically expanded to match the number of num_prompts, e.g., [1,1,1]. If an array is used, its length must match the number of num_prompts.                          |
 | max_concurrency | int / array | No       | 1, [1, 2, 3]    | Maximum concurrent in-flight requests. Same array / expansion rules as `request_rate` (mutually exclusive with QPS mode).                                                                                                                                                                                                             |
@@ -12,12 +12,37 @@ When running test cases, you may need to install the following dependencies:

 ```bash
 uv pip install ".[dev]"
-apt-get install -y ffmpeg
+apt-get install -y espeak-ng jq
 ```

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

+### CI job runners (L2–L4): logs, timing, and timeouts
+
+[`tools/run_ready_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/run_ready_jobs.sh), [`tools/run_merge_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/run_merge_jobs.sh), and [`tools/nightly/run_nightly_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/nightly/run_nightly_jobs.sh) share the same run-time behavior (via [`tools/run_jobs_common.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/run_jobs_common.sh)):
+
+**Log layout** (default under `logs/ready_jobs`, `logs/merge_jobs`, or `logs/nightly_jobs`; override with `--log-dir`):
+
+| Path | Description |
+|------|-------------|
+| `jobs/*.sh` | Generated per-step bash wrappers |
+| `jobs/.job_timeouts` | Job key → `timeout_in_minutes` from Buildkite YAML (when set) |
+| `<job_key>.log` | Tee output for each job |
+| `timing_summary.log` | Per-job duration and total wall time after the run completes |
+
+**Timing:** while each job runs, stderr prints `finished in …` or `failed after …`. When all jobs finish, a summary is printed and written to `timing_summary.log`, for example:
+
+```
+=== Job timing summary ===
+  TTS_VoxCPM2_Test  25m 40s  OK
+  Omni_Qwen3-Omni_Test  8m 05s  TIMED OUT
+Total wall time: 33m 45s (2 jobs)
+Failed jobs: 1/2
+```
+
+**Timeouts:** if a Buildkite step defines `timeout_in_minutes`, the runner wraps that job with `timeout ${N}m` (aligned with CI). Steps without `timeout_in_minutes` rely on any inline `timeout …` already present in the extracted pytest command. On timeout the process exits with code `124` and the summary marks the job as `TIMED OUT`.
+
 === "L1 level"

    ```bash
@@ -28,60 +53,114 @@ Our test scripts use the pytest framework. First, please use `git clone https://

 === "L2 level"

+    **Recommended:** run CI-aligned jobs from the repo root via [`tools/run_ready_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/run_ready_jobs.sh). The script reads [`.buildkite/test-ready.yml`](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-ready.yml), generates per-step bash wrappers, runs pytest, and tees logs under `logs/ready_jobs/` (see [CI job runners](#ci-job-runners-l2l4-logs-timing-and-timeouts) for timing and timeout behavior). Requires `bash`, `python3`, and PyYAML (`pip install pyyaml`).
+
+    ```bash
+    # All L2 and L1 jobs (default: every pytest step in test-ready.yml)
+    bash tools/run_ready_jobs.sh
+
+    # Skip L1-style "Simple Test" steps in test-ready.yml (group or "Simple ·" labels)
+    bash tools/run_ready_jobs.sh --skip-simple
+    bash tools/run_ready_jobs.sh --skip-simple --model-type diffusion
+
+    # Preview extracted commands without running
+    bash tools/run_ready_jobs.sh --dry-run
+
+    # Filter by model area (OR semantics): omni | tts | diffusion | all
+    bash tools/run_ready_jobs.sh --model-type omni
+    bash tools/run_ready_jobs.sh --model-type tts,diffusion
+
+    # Match a Buildkite label substring (e.g. one model suite)
+    bash tools/run_ready_jobs.sh --model-type tts --label-substr "VoxCPM2"
+
+    ```
+
+    **Ad hoc:** run a single test file or marker expression manually:
+
    ```bash
    cd tests
-    pytest -s -v -m "core_model and not cpu" --run-level=core_model
-    ```
-    If you only want to run a specific test case, you can use:
-    ```bash
    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
    ```
-    If you only want to run specific test cases on a particular platform, you can use:
-    ```bash
-    pytest -s -v -m "core_model and distributed_cuda and L4"  --run-level=core_model
-    ```
-    The latest test commands for various test suites can be found in the [pipeline](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-ready.yml).

 === "L3 level"

+    **Recommended:** run CI-aligned jobs from the repo root via [`tools/run_merge_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/run_merge_jobs.sh). The script reads [`.buildkite/test-merge.yml`](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-merge.yml), generates per-step bash wrappers, runs pytest, and tees logs under `logs/merge_jobs/` (see [CI job runners](#ci-job-runners-l2l4-logs-timing-and-timeouts)). Requires `bash`, `python3`, and PyYAML (`pip install pyyaml`).
+
    ```bash
-    pytest -s -v -m "advanced_model" --run-level=advanced_model
+    # All L3 and L1 jobs (default: every pytest step in test-merge.yml)
+    bash tools/run_merge_jobs.sh
+
+    # Skip L1-style "Simple Test" steps in test-merge.yml (group or "Simple ·" labels)
+    bash tools/run_merge_jobs.sh --skip-simple
+    bash tools/run_merge_jobs.sh --skip-simple --model-type omni
+
+    # Preview extracted commands without running
+    bash tools/run_merge_jobs.sh --dry-run
+
+    # Filter by model area (OR semantics): omni | tts | diffusion | all
+    bash tools/run_merge_jobs.sh --model-type diffusion
+    bash tools/run_merge_jobs.sh --model-type omni,tts
+
+    # Match a Buildkite label substring
+    bash tools/run_merge_jobs.sh --model-type diffusion --label-substr "Wan22"
+
    ```
-    If you only want to run a specific test case, you can use:
+
+    **Ad hoc:** run a single test file or marker expression manually:
+
    ```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
    ```
-    If you only want to run specific test cases on a particular platform, you can use:
-    ```bash
-    pytest -s -v -m "advanced_model and distributed_cuda and L4"  --run-level=advanced_model
-    ```
-    The latest L3 test commands for various test suites can be found in the [pipeline](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-merge.yml).


 === "L4 level"

+    **Recommended:** run CI-aligned nightly jobs from the repo root via [`tools/nightly/run_nightly_jobs.sh`](https://github.com/vllm-project/vllm-omni/blob/main/tools/nightly/run_nightly_jobs.sh). The script reads [`.buildkite/test-nightly.yml`](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-nightly.yml), generates per-step bash wrappers, runs pytest (perf jobs first, then aggregates perf JSON into Excel), and tees logs under `logs/nightly_jobs/` (see [CI job runners](#ci-job-runners-l2l4-logs-timing-and-timeouts); `generate_nightly_perf_excel` is included in the timing summary). Requires `bash`, `python3`, and PyYAML (`pip install pyyaml`).
+
+    ```bash
+    # All nightly jobs (default: test-type all, model-type all)
+    bash tools/nightly/run_nightly_jobs.sh
+
+    # Preview extracted commands without running
+    bash tools/nightly/run_nightly_jobs.sh --dry-run
+
+    # Test kind (--test-type, OR semantics): perf | acc | function | stability | local | all
+    bash tools/nightly/run_nightly_jobs.sh --test-type function
+    bash tools/nightly/run_nightly_jobs.sh --test-type perf,acc
+
+    # Model area (--model-type, OR semantics): omni | tts | diffusion | all
+    bash tools/nightly/run_nightly_jobs.sh --test-type function --model-type omni
+    bash tools/nightly/run_nightly_jobs.sh --test-type perf --model-type diffusion
+
+    # Stability scripts (tests/dfx/stability/scripts/) and local marker tests
+    bash tools/nightly/run_nightly_jobs.sh --test-type stability --model-type omni
+    bash tools/nightly/run_nightly_jobs.sh --test-type local --model-type tts
+    ```
+
+    Perf steps (label contains ``Perf Test``) run before other jobs; Excel output is written under ``logs/``. Diffusion nightly steps may also appear in [test-nightly-diffusion.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-nightly-diffusion.yml) on CI.
+
+    **Ad hoc:** run a single test file or marker expression manually:
+
    ```bash
    cd tests
-    pytest -s -v -m "full_model" --run-level=full_model
-    ```
-    If you only want to run a specific test case, you can use:
-    ```bash
    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 (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
    ```
-    Note: To run performance tests (defaults to ``test_qwen_omni.json``; use ``--test-config-file tests/dfx/perf/tests/test_tts.json`` for TTS):
+    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
-    pytest -s -v tests/dfx/perf/scripts/run_benchmark.py
-    ```
-    The latest L4 (nightly) test commands use the `full_model` marker and `--run-level full_model` (see [test-nightly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-nightly.yml) and [test-nightly-diffusion.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-nightly-diffusion.yml)). Example:
-
-    ```bash
-    cd tests
-    pytest -s -v -m "full_model and (omni or tts) and H100" --run-level=full_model
+    pytest -sv tests/dfx/perf/scripts/run_benchmark.py -m "full_model and tts and H100"
+    pytest -sv tests/dfx/perf/scripts/run_diffusion_benchmark.py -m "full_model and diffusion and H100"
+    pytest -sv tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json
+    pytest -sv tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
    ```
+    Nightly **Perf Test** jobs in ``test-nightly.yml`` use ``--test-config-file`` only (no ``-m``). E2e L4 function tests still use ``full_model`` + ``--run-level full_model`` (see [test-nightly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/test-nightly.yml)). Example:

 === "L5 level"

@@ -33,7 +33,8 @@ Defined in `pyproject.toml`:
 | `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                                           |
+| `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).

@@ -94,7 +95,7 @@ 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), NPU (A2/A3)
+- `res` must be a dict; supported resources: CUDA (L4/H100), ROCm (MI325), XPU (B60), MUSA (S5000), NPU (A2/A3)
 - `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:
@@ -120,6 +121,32 @@ def test_feature(omni_server):
    ...
 ```

+### JSON `mark` field (L4 perf configs)
+
+Perf JSON under `tests/dfx/perf/tests/` can attach pytest marks per **case** (`test_name` block). Parsed by `tests.dfx.conftest.resolve_pytest_marks` and applied to each parametrized run in `run_benchmark.py` / `run_diffusion_benchmark.py`.
+
+When `mark` is present, it must be an **array** with exactly one ``hardware_marks`` object (same semantics as `hardware_marks()` above), followed by registered pytest marker name strings such as `full_model`, `omni`, `tts`, or `diffusion`.
+
+```json
+{
+  "test_name": "test_cosmos3_t2i_official_demo_2gpu",
+  "mark": [
+    {"hardware_marks": {"res": {"cuda": "H100"}, "num_cards": 2}},
+    "full_model",
+    "diffusion"
+  ],
+  "server_type": "vllm-omni",
+  "server_params": { "...": "..." },
+  "benchmark_params": [{ "name": "1024x1024_steps4", "...": "..." }]
+}
+```
+
+- Local bulk load: `pytest -sv tests/dfx/perf/scripts/run_diffusion_benchmark.py -m "full_model and diffusion and H100"`
+- Nightly CI perf steps: `--test-config-file tests/dfx/perf/tests/test_<model>_vllm_omni.json` (file selects cases; no `-m`)
+- Result filenames use **runtime** GPU detection (`get_runtime_resource_label`); `H100` is omitted on the default CI pool
+
+See [L4 performance test examples](./test_examples/l4_performance_tests.inc.md) for the full schema.
+
 ## Add Support for a New Platform

 If you want to add support for a new platform (e.g., "tpu" for a new accelerator), follow these steps:
@@ -638,7 +638,7 @@ mkdir -p examples/online_serving/your_model_name

 - Script: `examples/offline_inference/your_model_name/end2end.py`
  - Parse args like BAGEL (`--model`, `--modality`, optional `--image-path`, `--steps`, etc.)
-  - Use `from vllm_omni.entrypoints.omni import Omni` (or `OmniDiffusion` if your model is diffusion-only)
+  - Use `from vllm_omni.entrypoints.omni import Omni` for both multi-stage and diffusion-only models
  - Save outputs (images/audio/video/text) with deterministic filenames (e.g., `output_0_0.png`)
 - Doc: `examples/offline_inference/your_model_name/README.md`
  - Include at least one runnable command, e.g.:
@@ -13,7 +13,7 @@ This document outlines the architectural design for vLLM-Omni.

 The primary goal of the vLLM-Omni project is to build the fastest and easiest-to-use open-source Omni-Modality model inference & serving engine. vLLM-Omni extends the original vLLM, which was created to support large language models for text-based autoregressive (AR) generation tasks. vLLM-Omni is designed to support:

-* **Non-textual Output:** Enables the integration, efficient processing and output of various data types, including but not limited to, images, audio, and video, alongside text.
+* **Non-textual Output:** Enables the integration, efficient processing and output of various data types, including but not limited to, images, audio, video, and action trajectories, alongside text.
 * **Non-Autoregressive Structure:** Support model structure beyond autoregressive, especially Diffusion Transformer (DiT), which is widely used in visual and audio generation.
 * **Integration with vLLM Core:** Maintain compatibility and leverage existing vLLM key modules and optimizations where applicable.
 * **Extensibility:** Design a modular and flexible architecture that can easily accommodate new modalities, model architectures, and output formats.
@@ -470,7 +470,7 @@ class Attention(nn.Module):
                 role="self", role_category=None, ...):
        # Resolve backend for this role from the active OmniDiffusionConfig
        config = get_current_diffusion_config_or_none()
-        attention_config = config.attention_config if config is not None else None
+        attention_config = config.diffusion_attention_config if config is not None else None
        attn_backend_cls, spec = get_attn_backend_for_role(
            role=role,
            head_size=head_size,
@@ -493,7 +493,7 @@ These backends provide the **kernel implementations** for attention computation.

 #### Backend Selection Mechanism

-Selection is driven by `AttentionConfig` on `OmniDiffusionConfig`, which carries a global `default` plus a `per_role` map of `AttentionSpec` entries. A role string is resolved in this order:
+Selection is driven by `OmniDiffusionConfig.diffusion_attention_config`, which carries a global `default` plus a `per_role` map of `AttentionSpec` entries. A role string is resolved in this order:

 ```python
 def get_attn_backend_for_role(role, head_size, attention_config=None, role_category=None):
@@ -909,7 +909,7 @@ def initialize_model_parallel(

 ```
 1. User Request
-   └─> OmniDiffusion.generate(prompt)
+   └─> Omni.generate(prompt)
       └─> Prepare OmniDiffusionRequest
           └─> DiffusionEngine.step(requests)

@@ -10,7 +10,7 @@ This guide demonstrates how to use the newly added features for extending vLLM-O
 Three main features enable custom pipeline extension:

 1. **`WorkerWrapperBase`**: A wrapper class that enables dynamic worker extension with custom functionality
-2. **`load_format`**: A parameter that controls how diffusion models are loaded, including support for custom pipelines
+2. **`diffusion_load_format`**: A parameter that controls how diffusion models are loaded, including support for custom pipelines
 3. **`CustomPipelineWorkerExtension`**: An extension class that enables pipeline re-initialization with custom implementations

 ## Features
@@ -27,15 +27,15 @@ Three main features enable custom pipeline extension:

 **Location:** `vllm_omni/diffusion/worker/diffusion_worker.py`

-### load_format Parameter
+### diffusion_load_format parameter

-The `load_format` parameter controls how diffusion models are loaded. It supports the following values:
+The `diffusion_load_format` parameter controls the initial diffusion model load. The relevant values are:

 - **`"default"`**: Standard model loading using the model registry (default behavior)
- **`"custom_pipeline"`**: Load a custom pipeline class specified by `custom_pipeline_name`
- **`"dummy"`**: Skip model loading (useful for testing or when pipeline will be initialized separately)
+- **`"dummy"`**: Skip the initial model load; use this with `custom_pipeline_args["pipeline_class"]` for the custom pipeline workflow below
+- **`"diffusers"`**: Load through the Hugging Face Diffusers adapter

-**Location:** `vllm_omni/diffusion/model_loader/diffusers_loader.py`
+**Location:** `vllm_omni/diffusion/data.py` (`OmniDiffusionConfig`)

 ### CustomPipelineWorkerExtension

@@ -34,7 +34,7 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr

 ### Pre-built wheels

-Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, 0.23.0, and 0.24.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
+Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, 0.23.0, 0.24.0, and 0.25.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).

 === "NVIDIA CUDA"

@@ -20,7 +20,7 @@ Therefore, it is recommended to install vLLM and vLLM-Omni with a **fresh new**

 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.24.0 --torch-backend=auto
+uv pip install vllm==0.25.0 --torch-backend=auto
 ```

 #### Installation of vLLM-Omni
@@ -39,13 +39,13 @@ uv pip install 'vllm-omni[demo]'
 # --8<-- [start:build-wheel-from-source]

 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.24.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.25.0 release version of the library

 ```bash
-uv pip install vllm==0.24.0 --torch-backend=auto
+uv pip install vllm==0.25.0 --torch-backend=auto
 ```

-The 0.24.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The 0.25.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.

 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -66,12 +66,12 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.24.0
+git checkout v0.25.0
 ```
 Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
 ```bash
-#For CUDA 13.0 (the default for v0.24.0; the wheel filename has no `+cu130` suffix)
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.24.0/vllm-0.24.0-cp38-abi3-manylinux_2_28_x86_64.whl
+#For CUDA 13.0 (the default for v0.25.0; the wheel filename has no `+cu130` suffix)
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.25.0/vllm-0.25.0-cp38-abi3-manylinux_2_28_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash
@@ -102,7 +102,7 @@ docker run --runtime nvidia --gpus 2 \
    --env "HF_TOKEN=$HF_TOKEN" \
    -p 8091:8091 \
    --ipc=host \
-    vllm/vllm-omni:v0.24.0 \
+    vllm/vllm-omni:v0.25.0 \
    vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```

@@ -13,7 +13,7 @@ vLLM-Omni current recommends the steps in under setup through Docker Images.

 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
+uv pip install vllm==0.25.0+rocm723 --extra-index-url https://wheels.vllm.ai/rocm/0.25.0/rocm723
 ```

 #### Installation of vLLM-Omni
@@ -34,13 +34,13 @@ uv pip install onnxruntime-rocm
 # --8<-- [start:build-wheel-from-source]

 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.24.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.25.0 release version of the library

 ```bash
-uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
+uv pip install vllm==0.25.0+rocm723 --extra-index-url https://wheels.vllm.ai/rocm/0.25.0/rocm723
 ```

-The pre-built 0.24.0 vLLM wheel targets ROCm 7.2.2. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The pre-built 0.25.0 vLLM wheel targets ROCm 7.2.3. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.

 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -58,7 +58,7 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.24.0
+git checkout v0.25.0
 python3 -m pip install -r requirements/rocm.txt
 python3 setup.py develop
 ```
@@ -130,7 +130,7 @@ docker run --rm \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --env "HF_TOKEN=$HF_TOKEN" \
  -p 8091:8091 \
-  vllm/vllm-omni-rocm:v0.24.0 \
+  vllm/vllm-omni-rocm:v0.25.0 \
  --model Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```

@@ -149,7 +149,7 @@ docker run --rm -it \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --env "HF_TOKEN=$HF_TOKEN" \
  --entrypoint bash \
-  vllm/vllm-omni-rocm:v0.24.0
+  vllm/vllm-omni-rocm:v0.25.0
 ```

 # --8<-- [end:pre-built-images]
@@ -9,8 +9,8 @@ The recommended way to use vLLM-Omni on NPU is through the vLLM-Ascend pre-built

 ```bash
 # vLLM-Ascend image
-# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.24.0
-# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.24.0-a3
+# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.25.0
+# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.25.0-a3
 docker run --rm \
    --name vllm-omni-npu \
    --shm-size=1g \
@@ -28,11 +28,11 @@ docker run --rm \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /root/.cache:/root/.cache \
    -p 8000:8000 \
-    -it quay.io/atlas-ci/vllm-ascend:0.24.0-a3 bash
+    -it quay.io/atlas-ci/vllm-ascend:0.25.0-a3 bash

 # Inside the container, install vLLM-Omni from source
 cd /vllm-workspace
-git clone -b v0.24.0 https://github.com/vllm-project/vllm-omni.git
+git clone -b v0.25.0 https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
 pip install -v -e . --no-build-isolation
 # or VLLM_OMNI_TARGET_DEVICE=npu pip install -v -e .
@@ -84,8 +84,8 @@ Supported images as following.
 Here's an example deployment command that has been verified on 4 x NPUs:

 ```bash
-# Atlas A2: quay.io/ascend/vllm-omni:v0.24.0
-# Atlas A3: quay.io/ascend/vllm-omni:v0.24.0-a3
+# Atlas A2: quay.io/ascend/vllm-omni:v0.25.0
+# Atlas A3: quay.io/ascend/vllm-omni:v0.25.0-a3
 docker run --rm \
    --name vllm-omni-a3 \
    --shm-size=4g \
@@ -110,7 +110,7 @@ docker run --rm \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v ~/.cache:/root/.cache \
    -p 8091:8091 \
-    -it quay.io/ascend/vllm-omni:v0.24.0-a3 bash
+    -it quay.io/ascend/vllm-omni:v0.25.0-a3 bash
 ```

 !!! tip
@@ -19,10 +19,10 @@ uv venv --python 3.12 --seed
 source .venv/bin/activate

 # On CUDA
-uv pip install vllm==0.24.0 --torch-backend=auto
+uv pip install vllm==0.25.0 --torch-backend=auto

 # On ROCm
-uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
+uv pip install vllm==0.25.0+rocm723 --extra-index-url https://wheels.vllm.ai/rocm/0.25.0/rocm723

 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
@@ -34,7 +34,7 @@ For additional installation methods — please see the [installation guide](inst
 !!! note
    It is important to install the same major & minor version of vLLM and vLLM Omni, otherwise things may not work as expected. If the versions are misaligned, you will see a warning when you import vLLM Omni.

-    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.24.0` and vLLM Omni `0.24.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
+    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.25.0` and vLLM Omni `0.25.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.

 ## Offline Inference

@@ -16,8 +16,35 @@ ROOT_DIR_RELATIVE = "../../../../.."
 EXAMPLE_DIR = ROOT_DIR / "examples"
 EXAMPLE_DOC_DIR = ROOT_DIR / "docs/user_guide/examples"
 NAV_FILE = ROOT_DIR / "docs/.nav.yml"
+MODEL_DISPLAY_NAMES_FILE = EXAMPLE_DIR / "model_display_names.yml"
 MAX_INLINE_MATERIAL_SIZE = 8 * 1024

+SERVING_MODE_TITLES = {
+    "offline_inference": "Offline inference",
+    "online_serving": "Online serving",
+}
+
+
+def load_model_display_names() -> dict[str, str]:
+    try:
+        with open(MODEL_DISPLAY_NAMES_FILE, encoding="utf-8") as f:
+            model_display_names = yaml.safe_load(f)
+    except (OSError, yaml.YAMLError) as exc:
+        raise ValueError(f"Failed to load MODEL_DISPLAY_NAMES_FILE at {MODEL_DISPLAY_NAMES_FILE}: {exc}") from exc
+
+    if not isinstance(model_display_names, dict) or not all(
+        isinstance(key, str) and isinstance(value, str) for key, value in model_display_names.items()
+    ):
+        raise ValueError(
+            "MODEL_DISPLAY_NAMES_FILE "
+            f"at {MODEL_DISPLAY_NAMES_FILE} must contain a YAML mapping "
+            "of string model IDs to string display names"
+        )
+    return model_display_names
+
+
+MODEL_DISPLAY_NAMES = load_model_display_names()
+

 def fix_case(text: str) -> str:
    subs = {
@@ -54,6 +81,7 @@ class Example:
        main_file (Path): The main file in the directory.
        other_files (list[Path]): list of other files in the directory.
        title (str): The title of the document.
+        nav_title (str): The concise title used in navigation.

    Methods:
        __post_init__(): Initializes the main_file, other_files, and title attributes.
@@ -78,6 +106,16 @@ class Example:
    def is_code(self) -> bool:
        return self.main_file.suffix != ".md"

+    @property
+    def model_display_name(self) -> str | None:
+        if self.category not in SERVING_MODE_TITLES:
+            return None
+        return MODEL_DISPLAY_NAMES.get(self.path.stem)
+
+    @property
+    def nav_title(self) -> str:
+        return self.model_display_name or self.title
+
    def determine_main_file(self) -> Path:
        """
        Determines the main file in the given path.
@@ -151,13 +189,25 @@ class Example:
        return [file for file in self.path.rglob("*") if is_other_file(file)]

    def determine_title(self) -> str:
+        model_display_name = self.model_display_name
        if not self.is_code:
            # Specify encoding for building on Windows
            with open(self.main_file, encoding="utf-8") as f:
                first_line = f.readline().strip()
            match = re.match(r"^#\s+(?P<title>.+)$", first_line)
+            if model_display_name:
+                expected_title = f"{model_display_name}: {SERVING_MODE_TITLES[self.category]}"
+                actual_title = match.group("title") if match else first_line
+                if actual_title != expected_title:
+                    raise ValueError(
+                        f"Model example title mismatch in {self.main_file}: "
+                        f"expected '# {expected_title}', got {first_line!r}"
+                    )
+                return expected_title
            if match:
                return match.group("title")
+        elif model_display_name:
+            raise ValueError(f"Mapped model example must use a Markdown README: {self.path}")
        return fix_case(self.path.stem.replace("_", " ").title())

    def fix_relative_links(self, content: str) -> str:
@@ -300,7 +350,7 @@ def update_nav_file(examples: list[Example]):
        for example in category_examples:
            doc_path = EXAMPLE_DOC_DIR / example.category / f"{example.path.stem}.md"
            rel_path = doc_path.relative_to(ROOT_DIR / "docs")
-            category_items.append({example.title: rel_path.as_posix()})
+            category_items.append({example.nav_title: rel_path.as_posix()})

        if category_items:
            # Format category name (e.g., "offline_inference" -> "Offline Inference")
@@ -34,6 +34,7 @@ th {
 | `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` | ✅︎ | | | |
 | `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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, V2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
@@ -66,7 +66,7 @@ Content-Type: application/json
 |-----------|------|---------|-------------|
 | `input` | string | **required** | Text prompt describing the audio to generate |
 | `model` | string | server's model | Model to use (optional, should match server if specified) |
-| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, aac, opus |
+| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, opus |
 | `speed` | float | 1.0 | Playback speed (0.25 - 4.0) |

 #### Diffusion Parameters
@@ -90,7 +90,6 @@ Returns binary audio data with the appropriate `Content-Type` header:
 | `mp3` | `audio/mpeg` |
 | `flac` | `audio/flac` |
 | `pcm` | `audio/pcm` |
-| `aac` | `audio/aac` |
 | `opus` | `audio/opus` |

 ## Examples
@@ -285,7 +284,7 @@ Pydantic validation failure (e.g. invalid `response_format`, `speed` out of rang
    "detail": [
        {
            "type": "literal_error",
-            "msg": "Input should be 'wav', 'pcm', 'flac', 'mp3', 'aac' or 'opus'",
+            "msg": "Input should be 'wav', 'pcm', 'flac', 'mp3' or 'opus'",
            ...
        }
    ]
@@ -252,7 +252,7 @@ Validation errors (missing required fields):

 ```bash
 # Check if server is responding
-curl -X http://localhost:8000/v1/images/edit \
+curl http://localhost:8000/v1/images/edits \
  -F "prompt='test'"
 ```

@@ -138,7 +138,7 @@ Content-Type: application/json
 | `model` | string | server's model | Model to use (optional, should match server if specified) |
 | `n` | integer | 1 | Number of images to generate (1-10) |
 | `size` | string | model defaults | Image dimensions in WxH format (e.g., "1024x1024", "512x512") |
-| `response_format` | string | "b64_json" | Response format (only "b64_json" supported) |
+| `response_format` | string | "b64_json" | Response format (`"b64_json"` or `"file"`) |
 | `user` | string | null | User identifier for tracking |

 #### vllm-omni Extension Parameters
@@ -109,7 +109,7 @@ Content-Type: application/json
 | `input` | string | **required** | The text to synthesize into speech |
 | `model` | string | server's model | Model to use (optional, should match server if specified) |
 | `voice` | string | "vivian" | Speaker name (e.g., vivian, ryan, aiden) |
-| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, aac, opus |
+| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, opus |
 | `speed` | float | 1.0 | Playback speed (0.25-4.0) |

 #### vLLM-Omni Extension Parameters
@@ -83,7 +83,7 @@ When constructing `OmniDiffusionConfig` directly:
 from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, OmniDiffusionConfig

 config = OmniDiffusionConfig(
-    attention_config=AttentionConfig(
+    diffusion_attention_config=AttentionConfig(
        default=AttentionSpec(backend="FLASH_ATTN"),
        per_role={
            "cross": AttentionSpec(backend="TORCH_SDPA"),
@@ -33,7 +33,7 @@ m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_cpu_offload=True)

 **CLI:**
 ```bash
-vllm-omni serve diffusion Wan-AI/Wan2.2-T2V-A14B-Diffusers --enable-cpu-offload
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --enable-cpu-offload
 ```

 ### To Support a Model
@@ -82,6 +82,33 @@ mutual exclusion: when one group runs, the other moves to CPU.
 - Support single GPU only for now


+### Component offloading for split models (e.g. Cosmos3)
+
+Some models split their transformer into mutually-exclusive *components* that run
+in different phases of a single forward pass rather than as separate pipeline
+components -- e.g. Cosmos3's understanding (reasoner) component runs once per
+generation while the generation (generator) component runs every denoising step.
+Such models have no separate text encoder to swap against, so the transformer
+owns a small model-local offload path and wraps each phase with
+`with self._offload_context(name):`
+
+```python
+class Cosmos3VFMTransformer(nn.Module):
+    def forward(self, ...):
+        with self._offload_context("reasoner"):
+            ...  # understanding pass, runs once
+        with self._offload_context("generator"):
+            ...  # denoising pass, runs every step
+```
+
+Model-level offloading then keeps exactly one component GPU-resident at a time
+(the other on CPU), reusing the same `SequentialOffloadHook` `.to()` movers. The
+pipeline opts in by exposing `enable_omni_model_cpu_offload` (which drives the
+transformer's `enable_model_cpu_offload` and pins the VAE). Layerwise offloading
+works for these models too -- each component declares its own block container via
+`_layerwise_offload_blocks_attrs`.
+
+
 ## Layerwise (Blockwise) Offloading

 ### How It Works
@@ -119,10 +146,10 @@ m = Omni(model="Wan-AI/Wan2.2-I2V-A14B-Diffusers", enable_layerwise_offload=True
 **CLI:**
 ```bash
 # Text-to-video
-vllm-omni serve diffusion Wan-AI/Wan2.2-T2V-A14B-Diffusers --enable-layerwise-offload
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --enable-layerwise-offload

 # Or image-to-video
-vllm-omni serve diffusion Wan-AI/Wan2.2-I2V-A14B-Diffusers --enable-layerwise-offload
+vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers --omni --enable-layerwise-offload
 ```

 ### To Support a Model
@@ -145,9 +172,9 @@ class Flux2Transformer2DModel(nn.Module):
 ```

 ### Limitations
- Cold start latency increases because of
-    1) components are loaded to CPU first at the very first during initialization,
-    2) weight consolidation and pinning
+- Cold start latency increases because offloaded components must be moved to CPU
+  during setup; layerwise offload may add extra weight consolidation and pinning
+  work.
 - Performance depends on compute cost and H2D bandwidth as well
 - Support single GPU only for now

@@ -178,11 +205,18 @@ Both strategies use vLLM-Omni's hook registry system (`HookRegistry` and `ModelH

 ```
 OffloadBackend (base class)
-├── ModelLevelOffloadBackend → uses SequentialOffloadHook
+├── ModelLevelOffloadBackend → uses SequentialOffloadHook (.to() swap)
+│                              (delegates to a pipeline's enable_omni_model_cpu_offload
+│                               for split models like Cosmos3)
 └── LayerWiseOffloadBackend → uses LayerwiseOffloadHook
 ```

-Factory function `get_offload_backend()` selects the appropriate backend based on configuration.
+Factory function `get_offload_backend()` selects the appropriate backend based on
+configuration.
+
+For split models, `ModelLevelOffloadBackend.enable()` detects a pipeline's
+`enable_omni_model_cpu_offload` hook and delegates to it; Cosmos3 then swaps its
+reasoner/generator components inside the model forward pass.


 ## Supported Models
@@ -47,7 +47,8 @@ For supported Wan2.2 pipelines, the execution order is:
 1. Diffusion worker finishes denoising and decodes the raw video tensor.
 2. Worker-side model-specific post-processing runs.
 3. If frame interpolation is enabled, RIFE interpolates the decoded video
-   tensor on the worker side and records a FPS multiplier in `custom_output`.
+   tensor on the worker side and records a FPS multiplier in
+   `metadata.video.video_fps_multiplier`.
 4. The API server receives the already-interpolated video and only performs
   MP4 export.

@@ -105,7 +105,7 @@ The following tables show which models support each feature:

 > Notes:

-> 1. CPU Offload has two methods: Module-wise (default for models with DiT + text encoder) and Layerwise. The tables below show **Layerwise support** only.
+> 1. CPU Offload has two methods: Module-wise (default for models with DiT + text encoder) and Layerwise. The tables below show **Layerwise support** only. Split models like Cosmos3 (no separate text encoder) swap their reasoner/generator components for module-wise offload; see the [CPU Offload Guide](diffusion/cpu_offload_diffusion.md).
 > 2. The **💾Quantization** column is collapsed for readability. See [Quantization Overview](quantization/overview.md) for per-method and per-model support details.

 ### ImageGen
@@ -121,6 +121,7 @@ The following tables show which models support each feature:
 | **GLM-Image**            |     ❌     |     ❌      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Hidream-I1-Full**        |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **HunyuanImage3**        |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ✅        |        ❌         |
+| **Krea 2**               |     ❌     |     ✅      |           ❌           |       ❌        |         ❌         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LongCat-Image**        |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **LongCat-Image-Edit**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **MagiHuman**            |     ❌     |     ❌      |           ❌           |       ❓        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -143,7 +144,8 @@ The following tables show which models support each feature:
 > Notes:
 > 1. Nextstep_1(T2I) does not support cache acceleration methods such as TeaCache or Cache-DiT.
 > 2. `Tongyi-MAI/Z-Image-Turbo` and `SII-GAIR/daVinci-MagiHuman-Base-1080p` are distilled models with minimal NFEs; CFG-Parallel is not necessary.
-> 3. Cosmos3 T2I uses `Cosmos3OmniDiffusersPipeline` with `modalities=["image"]`. Model-level CPU offload is not supported; use layerwise offload.
+> 3. Cosmos3 T2I uses `Cosmos3OmniDiffusersPipeline` with `modalities=["image"]`. Model-level CPU offload swaps the nested UND reasoner and GEN generator pathways; layerwise offload remains available for blockwise GEN/UND offload.
+> 4. Krea 2 currently supports single-GPU inference plus LoRA, Cache-DiT, HSDP, CPU/layerwise offload, and VAE-patch-parallel (decode). TP/SP/CFG-Parallel are not yet wired. The few-step distilled (Turbo) checkpoint uses `is_distilled=true` (fixed timestep shift `mu=1.15`); generate at 2048x2048 by default with `num_inference_steps≈8` and `guidance_scale=0`. The Raw checkpoint uses 1024x1024, `num_inference_steps=28`, and `guidance_scale=4.5`.

 ### VideoGen

@@ -257,7 +257,7 @@ python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
 | :------- | :--- | :------ | :---------- |
 | `--deploy-config` | string | `None` | Path to deploy YAML (auto-detected if omitted) |
 | `--stage-configs-path` | string | `None` | [Deprecated] Legacy path to `stage_args` YAML; prefer `--deploy-config` |
-| `--worker-backend` | choice | `process` | `process` or `ray` |
+| `--worker-backend` | choice | `multi_process` | `multi_process` or `ray` |
 | `--ray-address` | string | `None` | Ray cluster address |
 | `--quantization` | string | `None` | Quantization method (e.g. `fp8`) |
 | `--log-stats` | flag | `False` | Enable statistics logging |
@@ -26,18 +26,22 @@ Stage 0 (AR Model)                Stage 1 (Diffusion)

 ```python
 from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams

 if __name__ == "__main__":
    omni = Omni(model="zai-org/GLM-Image")
    outputs = omni.generate(
        "A photorealistic mountain landscape at sunset",
-        sampling_params={
-            "height": 1024,
-            "width": 1024,
-            "num_inference_steps": 50,
-            "guidance_scale": 1.5,
-            "seed": 42,
-        },
+        sampling_params_list=[
+            omni.default_sampling_params_list[0],
+            OmniDiffusionSamplingParams(
+                height=1024,
+                width=1024,
+                num_inference_steps=50,
+                guidance_scale=1.5,
+                seed=42,
+            ),
+        ],
    )
    outputs[0].request_output.images[0].save("output.png")
 ```
@@ -46,6 +50,7 @@ if __name__ == "__main__":

 ```python
 from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams

 if __name__ == "__main__":
    omni = Omni(model="zai-org/GLM-Image")
@@ -56,13 +61,16 @@ if __name__ == "__main__":
                "image": "input.png",
            },
        },
-        sampling_params={
-            "height": 1024,
-            "width": 1024,
-            "num_inference_steps": 50,
-            "guidance_scale": 1.5,
-            "seed": 42,
-        },
+        sampling_params_list=[
+            omni.default_sampling_params_list[0],
+            OmniDiffusionSamplingParams(
+                height=1024,
+                width=1024,
+                num_inference_steps=50,
+                guidance_scale=1.5,
+                seed=42,
+            ),
+        ],
    )
    outputs[0].request_output.images[0].save("output.png")
 ```
@@ -1,4 +1,4 @@
-# MiMo-Audio Offline Inference
+# MiMo-Audio: Offline inference

 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/mimo_audio>.

@@ -1,4 +1,4 @@
-# Qwen2.5-Omni
+# Qwen2.5-Omni: Offline inference

 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/qwen2_5_omni>.

@@ -1,4 +1,4 @@
-# Qwen3-Omni
+# Qwen3-Omni: Offline inference

 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/qwen3_omni>.

@@ -1,4 +1,4 @@
-# AURA Omni Pipeline
+# AURA Omni: Online serving

 `aura_omni` wires ASR, AURA, and Qwen3-TTS into one vLLM-Omni pipeline:

@@ -1,4 +1,4 @@
-# Online serving Example of vLLM-Omni for MiMo-Audio
+# MiMo-Audio: Online serving

 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/mimo_audio>.

@@ -1,4 +1,4 @@
-# Qwen2.5-Omni
+# Qwen2.5-Omni: Online serving

 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/qwen2_5_omni>.

@@ -1,4 +1,4 @@
-# Qwen3-Omni
+# Qwen3-Omni: Online serving

 Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/qwen3_omni>.

@@ -139,7 +139,7 @@ Content-Type: application/json
 |-----------|------|---------|-------------|
 | `input` | string | **required** | Text prompt describing the audio to generate |
 | `model` | string | server's model | Model to use (optional, should match server if specified) |
-| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, aac, opus |
+| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, opus |
 | `speed` | float | 1.0 | Playback speed (0.25 - 4.0) |
 | `audio_length` | float | null | Audio duration in seconds (max ~47s for `stable-audio-open-1.0`) |
 | `audio_start` | float | 0.0 | Audio start time in seconds |
@@ -3,7 +3,7 @@
 ## Overview

 [AutoRound](https://github.com/intel/auto-round) produces pre-quantized
-checkpoints for LLMs, VLMs, and diffusion models. vLLM-Omni reads the
+checkpoints for LLMs, VLMs, diffusion models, and world models. vLLM-Omni reads the
 checkpoint's `config.json` and auto-detects
 `quantization_config.quant_method = "auto-round"`.

@@ -36,6 +36,13 @@ guide. AutoRound is Intel-supported.
 | Wan2.2-T2V | `Intel/Wan2.2-T2V-A14B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
 | Wan2.2-TI2V | `Intel/Wan2.2-TI2V-5B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |

+### World Model (Cosmos3)
+
+| Model | Checkpoint | Scope | Scheme | Backend |
+|-------|------------|-------|--------|---------|
+| Cosmos3-Nano | `Intel/Cosmos3-Nano-int4-AutoRound` | World-model transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
+| Cosmos3-Super | `Intel/Cosmos3-Super-int4-AutoRound` | World-model transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
+
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)

 | Model | Checkpoint | Scope | Scheme | Backend |
@@ -0,0 +1,120 @@
+# BitsAndBytes W4 Quantization
+
+## Overview
+
+BitsAndBytes 4-bit quantization supports weight-only NF4/FP4 diffusion transformer
+inference on CUDA GPUs. It quantizes BF16/FP16 weights at load time; activations
+stay in BF16/FP16. No pre-quantized checkpoint is required.
+
+This is an online (dynamic) path: load the normal HuggingFace checkpoint and
+quantize during model loading via the `bitsandbytes` CUDA kernels.
+
+## Hardware Support
+
+| Device | Support |
+|--------|---------|
+| NVIDIA CUDA GPU (SM 75+) | ✅ |
+| NVIDIA Blackwell GPU (SM 100+) | ✅ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ |
+| NVIDIA Ampere GPU (SM 80+) | ✅ |
+| AMD ROCm | ❌ |
+| Intel XPU | ❌ |
+| Ascend NPU | ❌ |
+
+Legend: `✅` supported, `❌` unsupported. Non-CUDA platforms raise in
+`get_quant_method()`; CUDA requires compute capability 7.5+ (`SM 75+`).
+
+Requires the optional `bitsandbytes` package (`pip install bitsandbytes`).
+
+## Model Type Support
+
+### Diffusion Model (Qwen-Image, Wan2.2)
+
+| Model | HF models | CUDA | Mode | Recommendation |
+|-------|-----------|:----:|------|----------------|
+| Z-Image | `Tongyi-MAI/Z-Image-Turbo` | Yes | Online W4 weight-only | All heavy linear layers; sensitive embedders stay BF16 |
+| Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | Not validated | Online W4 weight-only | Compare vs BF16 before enabling |
+| Wan2.2 | Wan2.2 diffusion pipelines | Not validated | Online W4 weight-only | Validate before enabling in docs |
+
+Other diffusion models may work if their transformer uses supported linear
+layers, but they are not validated in this guide.
+
+### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
+
+| Model | Scope | Status | Notes |
+|-------|-------|--------|-------|
+| Qwen3-Omni | Thinker language-model stage | Not validated | Prefer checkpoint-supported ModelOpt FP8 or AutoRound paths |
+| Qwen3-TTS | TTS language-model stage | Not validated | No BitsAndBytes TTS stage support is documented |
+
+### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
+
+| Model | Scope | Status | Notes |
+|-------|-------|--------|-------|
+| BAGEL | Stage-specific transformer or DiT module | Not validated | Requires explicit stage routing |
+| GLM-Image | Stage-specific transformer or DiT module | Not validated | Requires quality comparison with BF16 |
+
+## Configuration
+
+Python API:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(model="Tongyi-MAI/Z-Image-Turbo", quantization="bitsandbytes")
+
+omni_with_skips = Omni(
+    model="Tongyi-MAI/Z-Image-Turbo",
+    quantization_config={
+        "method": "bitsandbytes",
+        "ignored_layers": ["to_out"],
+    },
+)
+
+outputs = omni.generate(
+    "A cup of coffee on the table",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+CLI:
+
+```bash
+python text_to_image.py --model Tongyi-MAI/Z-Image-Turbo --quantization bitsandbytes
+python text_to_image.py --model Tongyi-MAI/Z-Image-Turbo --quantization bitsandbytes --ignored-layers "to_out"
+vllm serve Tongyi-MAI/Z-Image-Turbo --omni --quantization bitsandbytes
+```
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `method` | str | - | Quantization method (`"bitsandbytes"`) |
+| `quant_type` | str | `"nf4"` | 4-bit data type: `"nf4"` (recommended) or `"fp4"` |
+| `compress_statistics` | bool | `True` | Double-quantize block scaling statistics for better accuracy |
+| `ignored_layers` | list[str] | `[]` | Layer name patterns to keep in BF16/FP16 |
+
+## Validation and Notes
+
+On Z-Image-Turbo (single GPU, 1024×1024, 50 steps), BitsAndBytes W4 typically
+reduces peak VRAM from roughly 24.5 GiB (BF16) to roughly 17 GiB. Compare output
+quality against a BF16 baseline before enabling on new models.
+
+Multi-GPU tensor parallelism (`tensor_parallel_size` > 1) is not validated for
+BitsAndBytes in diffusion models; each rank quantizes its own weight shard
+independently.
+
+If quality regresses, use `ignored_layers` to keep sensitive projections in BF16
+(for example `to_out` or `w2`).
+
+```python
+omni = Omni(
+    model="Tongyi-MAI/Z-Image-Turbo",
+    quantization="bitsandbytes",
+    cache_backend="tea_cache",
+    cache_config={"rel_l1_thresh": 0.2},
+)
+```
+
+Only add a new model to the supported table after comparing the BitsAndBytes
+output against a BF16 baseline and documenting any required `ignored_layers`.
@@ -0,0 +1,12 @@
+# Canonical display names for model examples that have both offline and online pages.
+# The docs generator combines these names with the serving mode for each page H1.
+aura_omni: AURA Omni
+covo_audio: Covo-Audio-Chat
+dynin_omni: Dynin-Omni
+lance: Lance
+mimo_audio: MiMo-Audio
+ming_flash_omni: Ming-flash-omni 2.0
+qwen2_5_omni: Qwen2.5-Omni
+qwen3_omni: Qwen3-Omni
+sensenova_u1: SenseNova-U1
+step_audio2: Step-Audio2
@@ -1,4 +1,4 @@
-# AURA Omni
+# AURA Omni: Offline inference

 This example runs the native AURA Omni pipeline offline:

@@ -1,4 +1,4 @@
-# Covo-Audio-Chat (Offline Inference)
+# Covo-Audio-Chat: Offline inference

 ## Setup

@@ -1,4 +1,4 @@
-# Dynin-Omni Offline End2End Example
+# Dynin-Omni: Offline inference

 This folder contains a unified offline inference entrypoint:

@@ -1,4 +1,4 @@
-# Lance (ByteDance) — offline inference
+# Lance: Offline inference

 [Lance](https://huggingface.co/bytedance-research/Lance) is a 3B unified
 autoregressive + diffusion multimodal model on a Qwen2.5-VL backbone. It is
@@ -293,16 +293,19 @@ def main():

    if args.modality in ("img2text", "video2text"):
        for i, req_output in enumerate(outputs):
-            text = getattr(req_output, "output", None) or getattr(req_output, "text", None)
-            if text is None and hasattr(req_output, "custom_output"):
-                text = (req_output.custom_output or {}).get("text_output")
+            multimodal_output = getattr(req_output, "multimodal_output", {}) or {}
+            metadata = multimodal_output.get("metadata", {}) if isinstance(multimodal_output, dict) else {}
+            text_metadata = metadata.get("text", {}) if isinstance(metadata, dict) else {}
+            text = (multimodal_output.get("text") if isinstance(multimodal_output, dict) else None) or (
+                text_metadata.get("text_output") if isinstance(text_metadata, dict) else None
+            )
+            text = text or getattr(req_output, "output", None) or getattr(req_output, "text", None)
            print(f"[Output {i}] {text}")
        return

    if args.modality in ("text2video", "video2video", "image2video"):
        for i, req_output in enumerate(outputs):
-            custom = getattr(req_output, "custom_output", None) or {}
-            frames = custom.get("video_frames")
+            frames = getattr(req_output, "images", None)
            if not frames:
                print(f"[Output {i}] no video frames returned")
                continue
@@ -835,12 +835,15 @@ def run_task(
    print(f"[unified] task={task} {width}x{height} frames={num_frames} cfg={cfg} steps={steps} seed={seed}", flush=True)

    outputs = list(omni.generate(prompts=formatted, sampling_params_list=params))
-    custom = getattr(outputs[0], "custom_output", None) or {}
-    # Lance/Bagel return text at outputs[0].output (string) or
-    # custom_output["text_output"].
+    multimodal_output = getattr(outputs[0], "multimodal_output", {}) or {}
+    metadata = multimodal_output.get("metadata", {}) if isinstance(multimodal_output, dict) else {}
+    text_metadata = metadata.get("text", {}) if isinstance(metadata, dict) else {}
    text_out = None
    if task in {"x2t_image", "x2t_video"}:
-        text_out = custom.get("text_output") or getattr(outputs[0], "output", None)
+        text_out = (multimodal_output.get("text") if isinstance(multimodal_output, dict) else None) or (
+            text_metadata.get("text_output") if isinstance(text_metadata, dict) else None
+        )
+        text_out = text_out or getattr(outputs[0], "output", None)
        if not isinstance(text_out, str):
            text_out = None

@@ -856,13 +859,11 @@ def run_task(
            gr.update(value=str(text_out or "(empty)"), visible=True),
        )
    if task in {"t2v", "i2v", "video_edit"}:
-        frames = custom.get("video_frames")
+        frames = getattr(outputs[0], "images", None)
        if frames is None:
            raise gr.Error("No video frames in output.")
        return (gr.update(visible=False), gr.update(value=_save_video(frames), visible=True), gr.update(visible=False))
-    # image tasks — orchestrator strips ``outputs[0].output`` across the
-    # IPC boundary but preserves ``custom_output``, so prefer the latter.
-    img_out = custom.get("image") or getattr(outputs[0], "output", None)
+    img_out = (getattr(outputs[0], "images", None) or [None])[0] or getattr(outputs[0], "output", None)
    if img_out is None:
        raise gr.Error("No image in output.")
    return (gr.update(value=_save_image(img_out), visible=True), gr.update(visible=False), gr.update(visible=False))
@@ -1,4 +1,4 @@
-# MiMo-Audio Offline Inference
+# MiMo-Audio: Offline inference

 This directory contains an offline demo for running MiMo-Audio models with vLLM Omni. It builds task-specific inputs and generates WAV files or text outputs locally.

@@ -1,4 +1,4 @@
-# Ming-flash-omni 2.0
+# Ming-flash-omni 2.0: Offline inference

 [Ming-flash-omni-2.0](https://github.com/inclusionAI/Ming) is an omni-modal model supporting text, image, video, and audio understanding, with text and speech outputs.

@@ -1,4 +1,4 @@
-# Qwen2.5-Omni
+# Qwen2.5-Omni: Offline inference

 ## Setup
 Please refer to the [stage configuration documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/) to configure memory allocation appropriately for your hardware setup.
@@ -1,4 +1,4 @@
-# Qwen3-Omni
+# Qwen3-Omni: Offline inference

 ## Setup
 Please refer to the [stage configuration documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/) to configure memory allocation appropriately for your hardware setup.
@@ -1,4 +1,4 @@
-# SenseNova-U1-8B-MoT
+# SenseNova-U1: Offline inference

 ## Architecture

@@ -278,12 +278,17 @@ def main():
    )

    for req_output in outputs:
-        custom = getattr(req_output, "_custom_output", {}) or {}
-        if args.print_think and custom.get("think_text"):
-            print(f"[Think]\n{custom['think_text']}\n")
+        multimodal_output = getattr(req_output, "multimodal_output", {}) or {}
+        metadata = multimodal_output.get("metadata", {}) if isinstance(multimodal_output, dict) else {}
+        text_metadata = metadata.get("text", {}) if isinstance(metadata, dict) else {}
+        think_text = text_metadata.get("think_text") if isinstance(text_metadata, dict) else None
+        if args.print_think and think_text:
+            print(f"[Think]\n{think_text}\n")

        if is_text_output:
-            text = custom.get("text_output", "")
+            text = multimodal_output.get("text", "") if isinstance(multimodal_output, dict) else ""
+            if not text:
+                text = text_metadata.get("text_output", "") if isinstance(text_metadata, dict) else ""
            if not text:
                text = getattr(req_output, "text", "") or ""
            print(f"[Response]\n{text}")
@@ -1,4 +1,4 @@
-# Step-Audio2 Offline Inference Examples
+# Step-Audio2: Offline inference

 This directory contains examples for running offline inference with Step-Audio2 using vLLM-Omni.

@@ -36,6 +36,7 @@ This folder provides several entrypoints for experimenting with text-to-image di
 | `black-forest-labs/FLUX.2-dev` | 1024 x 1024 | 65.7 | >80 (CPU offload required) |
 | `HunyuanImage-3.0` | 1024 x 1024 | 80.0 (TP≥3)  | 160 |
 | `HiDream-I1-Full` | 1024 x 1024 | 63.7  | 57.7 |
+| `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | 1024 (Raw) / 2048 (Turbo) | — | ~30 |

 !!! info
 *Peak VRAM:  based on basic single-card usage, batch size =1, without any acceleration/optimization features. FLUX.2-dev requires `--enable-cpu-offload` on a single 80 GiB GPU.
@@ -196,6 +197,37 @@ python examples/offline_inference/text_to_image/text_to_image.py \
  --output /output.png
 ```

+### Krea 2
+
+Krea 2 is published as diffusers-format checkpoints: `krea/Krea-2-Turbo` (few-step distilled) and `krea/Krea-2-Raw`.
+The pipeline class (`Krea2Pipeline`) is auto-detected from `model_index.json`. The distilled Turbo checkpoint sets
+`is_distilled=true` and is best run at 2048x2048 with few steps and guidance disabled:
+
+```bash
+# Few-step distilled (Turbo) checkpoint
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model krea/Krea-2-Turbo \
+  --prompt "a fox in the snow" \
+  --num-inference-steps 8 \
+  --guidance-scale 0.0 \
+  --height 2048 --width 2048 \
+  --output krea2.png
+```
+
+For the Raw checkpoint, use more steps with CFG enabled:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model krea/Krea-2-Raw \
+  --prompt "a fox in the snow" \
+  --num-inference-steps 28 \
+  --guidance-scale 4.5 \
+  --output krea2.png
+```
+
+`guidance-scale` follows the Krea 2 convention (`velocity = cond + guidance_scale * (cond - uncond)`); CFG is enabled
+whenever `guidance-scale > 0`.
+
 ### Batch Requests (Multiple Prompts)

 You can pass multiple prompts in a single `Omni.generate` call. `Omni`
@@ -202,9 +202,10 @@ def parse_args() -> argparse.Namespace:
        "--quantization",
        type=str,
        default=None,
-        choices=["fp8", "int8"],
+        choices=["fp8", "int8", "bitsandbytes"],
        help="Quantization method for the transformer. "
-        "Options: 'fp8' (FP8 W8A8 on Ada/Hopper, weight-only on older GPUs), 'int8' (Int8 W8A8). "
+        "Options: 'fp8' (FP8 W8A8 on Ada/Hopper, weight-only on older GPUs), "
+        "'int8' (Int8 W8A8), 'bitsandbytes' (NF4 4-bit weight-only, CUDA SM 75+). "
        "Default: None (no quantization, uses BF16).",
    )
    parser.add_argument(
@@ -1,4 +1,4 @@
-# AURA Omni Native Pipeline
+# AURA Omni: Online serving

 `aura_omni` serves AURA as a native multi-stage vLLM-Omni pipeline:

@@ -1,4 +1,4 @@
-# Covo-Audio-Chat
+# Covo-Audio-Chat: Online serving

 ## Installation

@@ -1,4 +1,4 @@
-# Dynin-Omni Online Serving Example
+# Dynin-Omni: Online serving

 ## Installation

@@ -1,4 +1,4 @@
-# Lance — online serving
+# Lance: Online serving

 OpenAI-compatible chat completions API for Lance, served by `vllm-omni`.

@@ -1,4 +1,4 @@
-# Online serving Example of vLLM-Omni for MiMo-Audio
+# MiMo-Audio: Online serving

 ## 🛠️ Installation

@@ -1,4 +1,4 @@
-# Ming-flash-omni 2.0
+# Ming-flash-omni 2.0: Online serving

 ## Installation

@@ -1,4 +1,4 @@
-# Qwen2.5-Omni
+# Qwen2.5-Omni: Online serving

 ## 🛠️ Installation

@@ -1,4 +1,4 @@
-# Qwen3-Omni
+# Qwen3-Omni: Online serving

 ## 🛠️ Installation

@@ -0,0 +1,91 @@
+# Replica Data Parallelism for Video Generation
+
+A runnable recipe + benchmark for **replica data parallelism** (replica-DP) on a
+video DiT. Replica-DP runs `N` independent diffusion engine replicas, one per GPU,
+and routes each request to a single replica. It scales **request throughput**
+near-linearly; it does **not** change single-request latency (for that, use an
+intra-request axis such as tensor/sequence/CFG parallelism).
+
+This example uses `Wan-AI/Wan2.2-TI2V-5B-Diffusers` with one GPU per replica.
+
+## Files
+
+| File | Purpose |
+|------|---------|
+| `wan2_2_ti2v_dp.yaml` | Diffusion stage config; `runtime.num_replicas` + `runtime.devices` drive the fan-out. |
+| `run_server.sh` | Substitutes `NUM_REPLICAS` / `DEVICES` into the config and serves. |
+| `bench_replica_dp.py` | Replica-agnostic load driver; reports throughput, latency, and a per-input isolation check. |
+
+## Run
+
+Start the server with the replica count you want (one GPU per replica):
+
+```bash
+# baseline
+NUM_REPLICAS=1 DEVICES=0        ./run_server.sh
+# 2 replicas
+NUM_REPLICAS=2 DEVICES=0,1      ./run_server.sh
+# 4 replicas
+NUM_REPLICAS=4 DEVICES=0,1,2,3  ./run_server.sh
+```
+
+In another shell, drive a fixed workload and read the throughput:
+
+```bash
+python bench_replica_dp.py --url http://127.0.0.1:8098 \
+    --num-requests 28 --concurrency 8 --label "replicas=4"
+```
+
+Re-run the server at `NUM_REPLICAS=1,2,4` (same client workload each time) and
+compare `videos/min`. Pass `--baseline-tpm <N=1 value>` to print a scaling factor.
+
+### Isolation check
+
+Identical outputs under a *shared* prompt+seed prove nothing -- every request
+would be identical regardless of isolation. Instead each request gets a distinct
+input (a prompt from the pool + a unique seed `base_seed+i`), and we check that a
+request's output is unchanged whether it runs alone or concurrently amid other
+distinct requests. Capture a per-input baseline on `N=1`, then verify it under
+`N>=2` concurrency:
+
+```bash
+# 1) against NUM_REPLICAS=1 (serial), record each input's output hash
+python bench_replica_dp.py --num-requests 8 --concurrency 1 \
+    --write-baseline baseline.json
+
+# 2) against NUM_REPLICAS>=2, replay the SAME inputs concurrently and verify
+python bench_replica_dp.py --num-requests 8 --concurrency 8 \
+    --check-baseline baseline.json
+```
+
+Step 2 exits non-zero if any request's output differs from its own baseline (a
+mismatch means cross-request state bleed or misrouting). Keep `--num-requests`,
+`--base-seed`, and the prompt source identical between the two runs so the input
+sets match.
+
+## Measured scaling
+
+Wan2.2-TI2V-5B (832×480, 33 frames, 30 steps), 4× A800-80GB (NVLink), one GPU per replica:
+
+| Replicas | Throughput (videos/min) | Scaling | Efficiency |
+|----------|-------------------------|---------|------------|
+| 1 | 4.71 | 1.00× | — |
+| 2 | 9.20 | 1.95× | 98% |
+| 4 | 18.02 | 3.83× | 96% |
+
+Per-request latency stayed flat (~13–14 s) across all replica counts. Isolation
+is checked per-input via the baseline/verify protocol above: each request's
+output under `N>=2` concurrency must match its own single-replica baseline.
+
+## Notes on the config surface
+
+- `runtime.num_replicas: N` fans out `N` replicas; `runtime.devices` assigns their
+  GPUs. With `tensor_parallel_size = 1`, list one GPU per replica
+  (`num_replicas * tensor_parallel_size` entries, pool mode).
+- Replica fan-out here comes from the stage config, not a CLI flag. (The headless /
+  multi-runtime launch path uses the process-local `--omni-dp-size-local`, which
+  requires `--stage-id`.)
+- This recipe serves via `--stage-configs-path`. Expressing replica-DP under the
+  newer `--deploy-config` schema is part of the ongoing serving-config alignment
+  (see the diffusion parallelism guide and the `#4707` discussion); `num_replicas`
+  is accepted by both paths.
@@ -0,0 +1,241 @@
+#!/usr/bin/env python3
+"""Throughput-scaling + isolation benchmark for diffusion replica data parallelism.
+
+Sends `--num-requests` text-to-video requests (up to `--concurrency` in flight)
+to a running vLLM-Omni video server via `POST /v1/videos/sync` (blocks until the
+clip is produced) and reports:
+
+    completed / failed, makespan (wall clock), throughput (videos/min),
+    per-request latency p50/p90/max (nearest-rank), and -- when a baseline
+    manifest is supplied -- a per-request isolation check.
+
+Each request gets a *distinct* input (a prompt from the pool + a unique seed),
+so an output is a deterministic function of its own input and nothing else.
+
+Measuring replica scaling
+-------------------------
+This client is replica-agnostic. Vary the *server's* replica count
+(`runtime.num_replicas` in the stage config; see run_server.sh) and keep the
+client workload fixed, then compare throughput:
+
+    N=1 -> baseline
+    N=2 -> expect ~2x
+    N=4 -> expect ~4x   (near-linear is the goal)
+
+Pass the single-replica videos/min via `--baseline-tpm` to print a scaling
+factor for the current run.
+
+Isolation / correctness
+-----------------------
+Byte-identical outputs under a *shared* prompt+seed prove nothing -- every
+request would be identical regardless of isolation. Instead, give each request a
+distinct input and check that its output is unchanged whether it runs alone or
+concurrently amid other distinct requests:
+
+    1. Against an N=1 server, capture a per-input baseline:
+         bench_replica_dp.py --num-requests 8 --concurrency 1 \
+             --write-baseline baseline.json
+    2. Against an N>=2 server, replay the *same* inputs concurrently and verify
+       each output matches its own baseline:
+         bench_replica_dp.py --num-requests 8 --concurrency 8 \
+             --check-baseline baseline.json
+
+A mismatch means a request's output changed under concurrency -- cross-request
+state bleed or misrouting -- and the run exits non-zero.
+
+Requires: requests (`pip install requests`).
+"""
+
+import argparse
+import concurrent.futures as cf
+import hashlib
+import json
+import math
+import sys
+import time
+from pathlib import Path
+
+import requests
+
+# Distinct prompts, cycled across requests; combined with a per-request unique
+# seed this makes every request a distinct input (see input_key()).
+DEFAULT_PROMPTS = [
+    "A cat playing piano, cinematic, high detail",
+    "A drone shot over a snowy mountain range at sunrise",
+    "A red sports car driving through a neon-lit city at night",
+    "Ocean waves crashing on a rocky shore, slow motion",
+    "A hot air balloon floating over green rolling hills",
+    "Timelapse of clouds moving over a desert canyon at dusk",
+    "A steam train crossing a stone viaduct in autumn",
+    "A jellyfish drifting through deep blue water, backlit",
+]
+
+
+def input_key(prompt, seed, args):
+    """Stable identity of a request's input; matches baseline<->verify by content, not order."""
+    return "\x1f".join([prompt, str(seed), args.size, str(args.num_frames), str(args.steps)])
+
+
+def nearest_rank(xs, q):
+    """Nearest-rank percentile; q in [0, 1]. Correct for tiny samples (unlike int(q*n)-1)."""
+    if not xs:
+        return float("nan")
+    s = sorted(xs)
+    rank = max(1, math.ceil(q * len(s)))
+    return s[rank - 1]
+
+
+def one_request(args, spec, idx, save):
+    """Fire one /v1/videos/sync request; return (ok, latency, nbytes, md5, key, err)."""
+    prompt, seed = spec["prompt"], spec["seed"]
+    key = input_key(prompt, seed, args)
+    # multipart form fields match POST /v1/videos/sync in the omni API server
+    form = {
+        "prompt": (None, prompt),
+        "size": (None, args.size),  # e.g. "832x480"
+        "num_frames": (None, str(args.num_frames)),
+        "num_inference_steps": (None, str(args.steps)),
+        "seed": (None, str(seed)),  # per-request unique -> distinct, reproducible input
+    }
+    if args.model:
+        form["model"] = (None, args.model)
+    if args.negative_prompt:
+        form["negative_prompt"] = (None, args.negative_prompt)
+
+    t0 = time.perf_counter()
+    try:
+        r = requests.post(f"{args.url}/v1/videos/sync", files=form, timeout=args.timeout)
+        dt = time.perf_counter() - t0
+        if r.status_code != 200:
+            return (False, dt, 0, "", key, f"HTTP {r.status_code}: {r.text[:160]}")
+        body = r.content
+        md5 = hashlib.md5(body).hexdigest()
+        if save and args.save_dir:
+            Path(args.save_dir).mkdir(parents=True, exist_ok=True)
+            (Path(args.save_dir) / f"req{idx:03d}_seed{seed}.mp4").write_bytes(body)
+        return (True, dt, len(body), md5, key, "")
+    except Exception as e:  # noqa: BLE001 - report any client-side failure
+        return (False, time.perf_counter() - t0, 0, "", key, repr(e))
+
+
+def build_specs(args):
+    """Deterministic list of distinct (prompt, seed) inputs; identical across baseline/verify runs."""
+    prompts = DEFAULT_PROMPTS
+    if args.prompts_file:
+        prompts = [ln.strip() for ln in Path(args.prompts_file).read_text().splitlines() if ln.strip()]
+    return [{"prompt": prompts[i % len(prompts)], "seed": args.base_seed + i} for i in range(args.num_requests)]
+
+
+def run_batch(args, specs, save):
+    """Fire specs concurrently; return (results, makespan). results = list of one_request tuples."""
+    results = []
+    wall0 = time.perf_counter()
+    with cf.ThreadPoolExecutor(max_workers=args.concurrency) as ex:
+        futs = {ex.submit(one_request, args, spec, i, save): i for i, spec in enumerate(specs)}
+        for done, f in enumerate(cf.as_completed(futs), 1):
+            res = f.result()
+            results.append(res)
+            good, dt, nbytes, _md5, _key, err = res
+            print(f"  [{done}/{len(specs)}] {'ok' if good else 'FAIL'} {dt:6.1f}s {(nbytes / 1e6):5.1f}MB {err}")
+    return results, time.perf_counter() - wall0
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Diffusion replica-DP throughput + isolation benchmark")
+    ap.add_argument("--url", default="http://127.0.0.1:8098", help="server base URL (no path)")
+    ap.add_argument("--num-requests", type=int, default=16, help="total requests to send")
+    ap.add_argument("--concurrency", type=int, default=4, help="max in-flight (set >= replica count)")
+    ap.add_argument("--warmup", type=int, default=-1, help="warmup requests before timing (default: =concurrency)")
+    ap.add_argument("--base-seed", type=int, default=42, help="request i uses seed base_seed+i (distinct inputs)")
+    ap.add_argument("--prompts-file", default="", help="optional file, one prompt per line (else a built-in pool)")
+    ap.add_argument("--negative-prompt", default="")
+    ap.add_argument("--size", default="832x480", help="WIDTHxHEIGHT")
+    ap.add_argument("--num-frames", type=int, default=33)
+    ap.add_argument("--steps", type=int, default=30, help="num_inference_steps")
+    ap.add_argument("--model", default="", help="optional; defaults to the served model")
+    ap.add_argument("--timeout", type=float, default=1800, help="per-request timeout (s)")
+    ap.add_argument("--save-dir", default="", help="save returned clips")
+    ap.add_argument("--label", default="", help="run label, e.g. 'replicas=2'")
+    ap.add_argument(
+        "--baseline-tpm", type=float, default=0.0, help="single-replica videos/min, to print a scaling factor"
+    )
+    ap.add_argument("--write-baseline", default="", help="run against N=1: write per-input md5 manifest to this path")
+    ap.add_argument("--check-baseline", default="", help="run against N>=2: verify each output matches this manifest")
+    args = ap.parse_args()
+
+    if args.write_baseline and args.check_baseline:
+        ap.error("--write-baseline and --check-baseline are mutually exclusive")
+
+    specs = build_specs(args)
+    warmup = args.concurrency if args.warmup < 0 else args.warmup
+
+    print(
+        f"-> {args.url}/v1/videos/sync | {args.num_requests} reqs "
+        f"| concurrency {args.concurrency} | warmup {warmup} | {args.size} {args.num_frames}f {args.steps}steps "
+        f"| {args.label}"
+    )
+
+    # Warmup: absorb each replica's first-call compile/init cost so it does not
+    # skew the measured makespan. Results are discarded and never saved.
+    if warmup > 0:
+        print(f"warming up ({warmup} reqs, not timed)...")
+        run_batch(args, [specs[i % len(specs)] for i in range(warmup)], save=False)
+
+    results, makespan = run_batch(args, specs, save=True)
+
+    ok = sum(1 for r in results if r[0])
+    fail = len(results) - ok
+    lat = [r[1] for r in results if r[0]]
+    errs = [r[5] for r in results if not r[0]]
+    got = {r[4]: r[3] for r in results if r[0]}  # input_key -> md5
+
+    print(f"\n===== result {args.label} =====")
+    print(f"completed {ok} / failed {fail} | makespan {makespan:.1f}s")
+    if ok:
+        tpm = ok / makespan * 60.0
+        print(f"throughput: {tpm:.2f} videos/min  ({ok / makespan:.4f} videos/s)")
+        print(
+            f"latency (s): p50={nearest_rank(lat, 0.50):.1f} "
+            f"p90={nearest_rank(lat, 0.90):.1f} "
+            f"min={min(lat):.1f} max={max(lat):.1f}"
+        )
+        if args.baseline_tpm > 0:
+            print(f"scaling vs baseline ({args.baseline_tpm:.2f}/min): {tpm / args.baseline_tpm:.2f}x")
+    if errs:
+        print("sample failures:", errs[:3])
+
+    # Isolation check: compare each request's output against its own single-replica baseline.
+    isolation_failed = False
+    if args.write_baseline:
+        Path(args.write_baseline).write_text(json.dumps(got, indent=2))
+        print(f"baseline: wrote {len(got)} per-input hashes -> {args.write_baseline}")
+    elif args.check_baseline:
+        baseline = json.loads(Path(args.check_baseline).read_text())
+        matched = mismatched = unknown = 0
+        for key, md5 in got.items():
+            if key not in baseline:
+                unknown += 1
+                print(f"  isolation: NO BASELINE for input {key!r} (not in the baseline manifest)")
+            elif baseline[key] == md5:
+                matched += 1
+            else:
+                mismatched += 1
+                print(f"  isolation: MISMATCH for input {key!r} (output changed under concurrency)")
+        # PASS requires the run to cover the *whole* baseline: a subset (fewer requests, or
+        # some failed) must not pass just because the inputs it did run happened to match.
+        unverified = sorted(set(baseline) - set(got))
+        for key in unverified:
+            print(f"  isolation: NOT VERIFIED for baseline input {key!r} (absent from this run)")
+        isolation_failed = mismatched > 0 or unknown > 0 or len(unverified) > 0
+        verdict = "PASS (isolated)" if not isolation_failed else "FAIL"
+        print(
+            f"isolation: {matched} matched / {mismatched} mismatched / {unknown} unknown / "
+            f"{len(unverified)} unverified vs baseline({len(baseline)}) -> {verdict}"
+        )
+
+    if fail > 0 or ok == 0 or isolation_failed:
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+# Serve N independent Wan2.2 TI2V-5B replicas (replica data parallelism).
+#
+# Usage:
+#   NUM_REPLICAS=1 DEVICES=0        ./run_server.sh      # baseline
+#   NUM_REPLICAS=2 DEVICES=0,1      ./run_server.sh
+#   NUM_REPLICAS=4 DEVICES=0,1,2,3  ./run_server.sh
+#
+# Then drive load with bench_replica_dp.py (see README.md) and compare
+# throughput across replica counts.
+set -euo pipefail
+
+MODEL="${MODEL:-Wan-AI/Wan2.2-TI2V-5B-Diffusers}"
+HOST="${HOST:-127.0.0.1}"
+PORT="${PORT:-8098}"
+NUM_REPLICAS="${NUM_REPLICAS:-1}"
+DEVICES="${DEVICES:-0}"   # one GPU per replica for tensor_parallel_size=1
+
+HERE="$(cd "$(dirname "$0")" && pwd)"
+TEMPLATE="$HERE/wan2_2_ti2v_dp.yaml"
+CFG="$(mktemp --suffix=_wan2_2_dp.yaml)"
+trap 'rm -f "$CFG"' EXIT
+
+sed -e "s/__NUM_REPLICAS__/${NUM_REPLICAS}/" \
+    -e "s/__DEVICES__/${DEVICES}/" \
+    "$TEMPLATE" > "$CFG"
+
+echo "Serving ${MODEL}: num_replicas=${NUM_REPLICAS} devices=${DEVICES} (config ${CFG})"
+
+vllm serve "$MODEL" --omni \
+    --host "$HOST" \
+    --port "$PORT" \
+    --stage-configs-path "$CFG"
@@ -0,0 +1,43 @@
+# Replica data-parallel recipe for Wan2.2 TI2V-5B (video DiT).
+#
+# Fans out `num_replicas` independent DiffusionEngine replicas, one per GPU,
+# and lets the serving layer route each request to a single replica. This
+# scales request throughput near-linearly; it does not change single-request
+# latency. See ./README.md for measured numbers and the `run_server.sh` driver.
+#
+# run_server.sh substitutes __NUM_REPLICAS__ / __DEVICES__ from env before serving.
+# For tensor_parallel_size = 1, `devices` lists one GPU per replica (pool mode):
+# num_replicas=4 -> devices "0,1,2,3".
+
+stage_args:
+  - stage_id: 0
+    stage_type: diffusion
+    runtime:
+      num_replicas: __NUM_REPLICAS__
+      devices: "__DEVICES__"
+      max_batch_size: 1
+    engine_args:
+      model_stage: dit
+      # `model_class_name` takes the diffusion registry KEY (WanPipeline), which
+      # resolves to the Wan22Pipeline class. Wan2.2 TI2V-5B is served by the base
+      # WanPipeline, which auto-detects TI2V-5B mode from model_index.json.
+      model_class_name: WanPipeline
+      max_num_seqs: 1
+      enforce_eager: true
+      trust_remote_code: true
+      distributed_executor_backend: "mp"
+      vae_use_tiling: true
+      parallel_config:
+        tensor_parallel_size: 1
+
+    final_output: true
+    final_output_type: video
+    is_comprehension: false
+    default_sampling_params:
+      seed: 42
+
+runtime:
+  enabled: true
+  defaults:
+    window_size: -1
+    max_inflight: 1
@@ -1,4 +1,4 @@
-# SenseNova-U1 Online Serving
+# SenseNova-U1: Online serving

 ## Launch the Server

@@ -1,4 +1,4 @@
-# Step-Audio2 Online Serving
+# Step-Audio2: Online serving

 This directory contains examples for running Step-Audio2 with vLLM-Omni's online serving API.

@@ -148,7 +148,7 @@ Do not copy latency or VRAM numbers between commits or machines. Record them
 only after running a latest-head sweep on the target hardware.

 For formal PR benchmarking, reuse
-`tests/dfx/perf/tests/test_ltx2_3_vllm_omni.json` with
-`tests/dfx/perf/scripts/run_diffusion_benchmark.py`. That config captures the
+`tests/dfx/perf/tests/test_ltx2_vllm_omni.json` with
+`tests/dfx/perf/scripts/run_diffusion_benchmark.py`. Each case in that file includes a JSON **`mark`** (`hardware_marks` + `full_model` + `diffusion`) for local `-m` filtering. That config captures the
 single-device eager baseline and CFG-parallel=2 case for a small 384x512,
 25-frame, 20-step workload.
@@ -44,6 +44,7 @@ recipes/
 | [`inclusionAI/Ming-flash-omni-2.0.md`](./inclusionAI/Ming-flash-omni-2.0.md) | Online serving for multimodal chat + standalone TTS | 4x H100 / 1x H100 80GB |
 | [`inclusionAI/Ming-omni-tts.md`](./inclusionAI/Ming-omni-tts.md) | Offline + online dense Ming TTS/audio generation | 1x H100 80GB / 1x AMD MI300X (ROCm 7.2) |
 | [`IndexTeam/IndexTTS-2.md`](./IndexTeam/IndexTTS-2.md) | Online serving for voice-cloned TTS with optional emotion control | 1x L4 24GB or larger CUDA GPU |
+| [`krea/Krea-2.md`](./krea/Krea-2.md) | Text-to-image (Turbo + Raw), offline + online, with LoRA | 1x H100 80GB |
 | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | Text-to-video and image-to-video serving | 1x H200 141GB |
 | [`LTX/LTX-2.3.md`](./LTX/LTX-2.3.md) | Text-to-video with audio generation (22B) | 1x GPU (96GB VRAM) |
 | [`MammothModa2/MammothModa2-Preview.md`](./MammothModa2/MammothModa2-Preview.md) | Text-to-image with the shared offline image example (AR → DiT) | 1x L40S 48GB / 1x ≥40GB GPU |
@@ -422,3 +422,185 @@ ffprobe -v error -show_entries stream=codec_type,nb_frames,width,height cosmos3_
  `generate_sound`/`sound_duration`, `guardrails`, `action_*`, ...) are declared
  once in `vllm_omni/model_extras/cosmos3.py` and forwarded through `--extra-body`;
  unknown keys for the model are dropped.
+
+## NPU
+
+### 1x Ascend 910B / 910C (Atlas A2 / A3) — Online serving
+
+#### Environment
+
+- OS: Linux (aarch64)
+- Python: 3.12+
+- Driver / runtime: CANN 8.5.1 + NNAL + Ascend 910B / 910C
+- vLLM version: match the repository requirements from your current checkout
+- vLLM-Ascend version: match the repository requirements from your current checkout
+- vLLM-Omni version or commit: use the commit you are deploying from
+
+#### Command
+
+Requires the `vllm-omni` package (or the `quay.io/atlas-ci/vllm-ascend` A2 / A3 container),
+which provides the `vllm serve … --omni` entrypoint used below.
+
+Safety guardrails are **on by default** (NVIDIA Open Model License). They load
+the **gated** `nvidia/Cosmos-1.0-Guardrail` model, so to keep them on you must:
+
+1. `pip install cosmos-guardrail`
+2. Accept the license at <https://huggingface.co/nvidia/Cosmos-1.0-Guardrail>
+3. Export a token with access: `export HF_TOKEN=hf_...`
+
+Then launch the recommended server:
+
+```bash
+vllm serve nvidia/Cosmos3-Nano \
+  --omni \
+  --host 0.0.0.0 --port 8000 \
+  --init-timeout 1800
+```
+
+To run **without** guardrails (you are responsible for license compliance),
+add `--no-guardrails` (no token/`cosmos-guardrail` needed). For tensor parallel
+add `--tensor-parallel-size 8`. `--quantization fp8` and
+`--enable-layerwise-offload` are not supported on NPU.
+The pipeline auto-resolves from `model_index.json`; pass
+`--model-class-name Cosmos3OmniDiffusersPipeline` to force it explicitly.
+
+#### Verification
+
+Best quality uses the JSON-upsampled prompts from `assets/` (download with
+`hf download nvidia/Cosmos3-Nano assets/ --local-dir Cosmos3-Nano`). Minimal
+self-contained examples:
+
+```bash
+curl http://localhost:8000/v1/models
+
+# Text-to-image -> /v1/images/generations  (1024x1024, 10 steps; base64 PNG)
+curl -sS -X POST http://localhost:8000/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "nvidia/Cosmos3-Nano",
+    "prompt": "A photorealistic red sports car on a city street at golden hour, cinematic lighting.",
+    "negative_prompt": "blurry, distorted, low quality",
+    "size": "1024x1024", "n": 1, "response_format": "b64_json",
+    "num_inference_steps": 10, "guidance_scale": 7.0, "seed": 42
+  }' | python -c "import sys,json,base64; open('cosmos3_t2i.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
+
+# Text-to-video -> /v1/videos/sync  (720p, 49 frames @ 24fps)
+curl -sS -X POST http://localhost:8000/v1/videos/sync \
+  -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Nano" \
+  -F "prompt=A robot arm is cleaning a plate in the kitchen" \
+  -F "negative_prompt=blurry, distorted, low quality, jittery, deformed" \
+  -F "size=1280x720" -F "num_frames=49" -F "fps=24" \
+  -F "num_inference_steps=20" -F "guidance_scale=6.0" \
+  -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F "seed=123" \
+  -o cosmos3_t2v.mp4
+
+# Image-to-video -> /v1/videos/sync with an uploaded reference image
+curl -sS -X POST http://localhost:8000/v1/videos/sync \
+  -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Nano" \
+  -F "prompt=The scene comes to life with smooth, natural motion." \
+  -F "size=1280x720" -F "num_frames=25" -F "fps=8" \
+  -F "num_inference_steps=10" -F "guidance_scale=6.0" \
+  -F "seed=42" \
+  -F "input_reference=@reference.jpg;type=image/jpeg" \
+  -o cosmos3_i2v.mp4
+
+# Video-to-video -> /v1/videos/sync with an uploaded reference video
+curl -sS -X POST http://localhost:8000/v1/videos/sync \
+  -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Nano" \
+  -F "prompt=Continue the same scene with smooth natural motion and consistent subjects." \
+  -F "size=1280x720" -F "num_frames=17" -F "fps=5" \
+  -F "num_inference_steps=10" -F "guidance_scale=6.0" \
+  -F "seed=42" \
+  -F "input_reference=@reference.mp4;type=video/mp4" \
+  -o cosmos3_v2v.mp4
+```
+
+#### Notes
+
+- **Measured latency (1x Ascend 910B / 910C, bf16, guardrails off):**
+  - T2I 1024² — 10 steps → ~8 s
+  - T2V 1280×720 @ 20 steps — 49 frames → ~55 s
+  - I2V 1280×720 @ 10 steps — 25 frames → ~25 s
+  - V2V 480×320 @ 10 steps — 17 frames → ~12 s
+- **Memory:** transformer ~17 GiB (bf16); peak ~46 GiB for 720p video on 1 NPU;
+  full repo (transformer + Wan VAE + Qwen3-VL vision encoder + audio tokenizer)
+  ~33 GB on disk.
+- **Determinism:** identical seed reproduces identical output on the same
+  hardware; outputs are not bit-identical across different GPU/NPU types.
+- **Supported sizes (per model card):** 256p / 480p / 720p at 16:9, 4:3, 1:1,
+  3:4, 9:16. Defaults: T2I 1024², 50 steps, guidance 7.0; T2V/I2V/V2V
+  1280×720, 35 steps, guidance 6.0, `flow_shift=10.0`.
+- **Key flags / params:** `--no-guardrails` (optional, to disable guardrails), `--init-timeout 1800`
+  (for model loading), `--tensor-parallel-size 8` for multi-NPU, and
+  `--model-class-name Cosmos3OmniDiffusersPipeline` to force the pipeline class.
+- **Known limitations:**
+  - Transfer V2V with `extra_params` (`edge`/`blur`/`depth`/`seg`/`wsm`) hits a
+    resolution-parsing bug; basic V2V without transfer hints works.
+  - FP8 online quantization and layerwise offload are not supported on NPU.
+
+### 1x Ascend 910B / 910C (Atlas A2 / A3) — Offline generation
+
+#### Environment
+
+- OS: Linux (aarch64)
+- Python: 3.12+
+- Driver / runtime: CANN 8.5.1 + NNAL + Ascend 910B / 910C
+- vLLM-Omni version or commit: use the commit you are deploying from
+
+#### Command
+
+The same offline task examples run on NPU; pass model-specific knobs via
+`--extra-body`. Guardrails are on by default — pass `"guardrails": false` for a
+quick local run (install `cosmos-guardrail` + accept the gated repo to enable them).
+
+```bash
+# Text-to-image -> examples/offline_inference/text_to_image
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model nvidia/Cosmos3-Nano \
+  --prompt "A photorealistic red sports car at golden hour, cinematic lighting." \
+  --negative-prompt "blurry, distorted, low quality" \
+  --height 1024 --width 1024 --num-inference-steps 50 --guidance-scale 7.0 \
+  --extra-body '{"flow_shift": 3.0, "guardrails": false}' \
+  --output cosmos3_t2i.png
+
+# Text-to-video -> examples/offline_inference/text_to_video
+python examples/offline_inference/text_to_video/text_to_video.py \
+  --model nvidia/Cosmos3-Nano \
+  --prompt "A robot arm is cleaning a plate in the kitchen." \
+  --negative-prompt "blurry, distorted, low quality, jittery, deformed" \
+  --height 720 --width 1280 --num-frames 189 --fps 24 \
+  --num-inference-steps 35 --guidance-scale 6.0 \
+  --extra-body '{"flow_shift": 10.0, "max_sequence_length": 4096, "guardrails": false,
+                 "use_resolution_template": false, "use_duration_template": false}' \
+  --output cosmos3_t2v.mp4
+
+# Image-to-video -> examples/offline_inference/image_to_video
+# (Cosmos3 bundles example frames under assets/; any RGB image works too.)
+python examples/offline_inference/image_to_video/image_to_video.py \
+  --model nvidia/Cosmos3-Nano \
+  --image /path/to/Cosmos3-Nano/assets/example_i2v_input.jpg \
+  --prompt "The scene comes to life with smooth, natural motion." \
+  --height 720 --width 1280 --num-frames 189 --fps 24 \
+  --num-inference-steps 35 --guidance-scale 6.0 \
+  --extra-body '{"flow_shift": 10.0, "max_sequence_length": 4096, "guardrails": false}' \
+  --output cosmos3_i2v.mp4
+```
+
+#### Verification
+
+```bash
+python -c "from PIL import Image; im=Image.open('cosmos3_t2i.png'); print('image', im.size, im.mode)"
+ffprobe -v error -show_entries stream=codec_type,nb_frames,width,height cosmos3_t2v.mp4
+```
+
+#### Notes
+
+- Guardrails are on by default on NPU with `cosmos-guardrail` installed. Pass
+  `"guardrails": false` in `--extra-body` to disable them (there is no
+  `--no-guardrails` flag for the offline scripts).
+- Video at the 189-frame default takes ~15 min/clip on 1 NPU; reduce
+  `--num-frames` for faster iteration.
@@ -0,0 +1,143 @@
+# Krea 2 for text-to-image (base + LoRA)
+
+## Summary
+
+- Vendor: Krea
+- Model: `krea/Krea-2-Turbo` (few-step distilled) and `krea/Krea-2-Raw`
+- Task: Text-to-image (T2I)
+- Mode: Offline engine and online OpenAI-compatible serving
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe for a known-good starting point to run the Krea 2 single-stream
+MMDiT with vLLM-Omni on a **single 80 GB NVIDIA H100** — both the offline
+`text_to_image.py` entrypoint and the online `/v1/images/generations` route,
+including PEFT LoRA adapters. `Krea2Pipeline` is auto-detected from
+`model_index.json`; guidance is checkpoint-aware (distilled Turbo runs no-CFG,
+Raw runs CFG).
+
+## References
+
+- Related example under `examples/`:
+  [`examples/offline_inference/text_to_image/README.md`](../../examples/offline_inference/text_to_image/README.md)
+  (see the "Krea 2" section)
+- Feature support:
+  [`docs/user_guide/diffusion_features.md`](../../docs/user_guide/diffusion_features.md)
+- Related PR: [#4730](https://github.com/vllm-project/vllm-omni/pull/4730)
+
+## Hardware Support
+
+This recipe documents a **single-GPU** CUDA layout on H100 80 GB. Add more
+platforms (ROCm / NPU) or multi-GPU layouts as community validation lands.
+
+## GPU
+
+### 1× H100 80GB
+
+#### Environment
+
+Versions from a working editable install (activate `vllm-omni/.venv` or your equivalent):
+
+- OS: Linux
+- Python: 3.12
+- Driver / runtime: NVIDIA driver **595.71.05** (CUDA 13.x)
+- torch: **2.11.0+cu130**
+- vLLM: **0.24.0**
+- vLLM-Omni: editable install from this repo (Git **`960ce8f6`**)
+- Transformers: **5.12.1**
+
+Krea 2 weights are ~30 GB and fit comfortably on a single 80 GB H100.
+
+#### Command
+
+Offline — few-step distilled (Turbo) checkpoint (2048×2048, no CFG):
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model krea/Krea-2-Turbo \
+  --prompt "a fox in the snow" \
+  --num-inference-steps 8 \
+  --guidance-scale 0.0 \
+  --height 2048 --width 2048 \
+  --output krea2_turbo.png
+```
+
+Offline — Raw checkpoint (1024×1024, CFG enabled):
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model krea/Krea-2-Raw \
+  --prompt "a fox in the snow" \
+  --num-inference-steps 28 \
+  --guidance-scale 4.5 \
+  --height 1024 --width 1024 \
+  --output krea2_raw.png
+```
+
+LoRA (offline) — pass a PEFT adapter with `--lora-path` / `--lora-scale`.
+`NagaSaiAbhinay/Krea-2-vllm-darkbrush-LoRA` is a vLLM-Omni-compatible repackaging
+of `krea/Krea-2-LoRA-darkbrush`; its trigger word is `monochrome ink wash style`:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model krea/Krea-2-Turbo \
+  --prompt "a fox in the snow, monochrome ink wash style" \
+  --num-inference-steps 8 --guidance-scale 0.0 \
+  --lora-path NagaSaiAbhinay/Krea-2-vllm-darkbrush-LoRA --lora-scale 1.0 \
+  --output krea2_darkbrush.png
+```
+
+Online — serve and query the OpenAI-compatible image route:
+
+```bash
+vllm serve krea/Krea-2-Turbo --omni --enforce-eager --port 8091
+```
+
+```bash
+curl -s http://localhost:8091/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "prompt": "a fox in the snow, monochrome ink wash style",
+    "size": "1024x1024",
+    "num_inference_steps": 8,
+    "guidance_scale": 0.0,
+    "seed": 42,
+    "lora": {"name": "darkbrush", "path": "NagaSaiAbhinay/Krea-2-vllm-darkbrush-LoRA", "scale": 1.0}
+  }' | jq -r '.data[0].b64_json' | base64 -d > krea2_online.png
+```
+
+Cache-DiT acceleration (offline) — add `--cache-backend cache_dit`:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model krea/Krea-2-Turbo --prompt "a fox in the snow" \
+  --num-inference-steps 8 --guidance-scale 0.0 \
+  --cache-backend cache_dit --output krea2_cachedit.png
+```
+
+#### Verification
+
+```bash
+# Offline: the script prints "Saved generated image to <path>"; confirm the file exists.
+ls -lh krea2_turbo.png
+```
+
+For the online route, decode the returned `b64_json` (see the `curl` above) and
+confirm a valid 1024×1024 PNG is written.
+
+#### Notes
+
+- Key flags: distilled Turbo → few steps (~8) with `guidance_scale=0.0`; Raw →
+  more steps (~28) with `guidance_scale>0` (CFG). `guidance-scale` follows the
+  Krea 2 convention `velocity = cond + guidance_scale * (cond - uncond)`.
+- LoRA: adapters must be PEFT format (`adapter_config.json` +
+  `adapter_model.safetensors`); the projections are `ReplicatedLinear`
+  (264 modules). Offline uses `--lora-path` (which sets
+  `OmniDiffusionSamplingParams.lora_request` / `lora_scale`); online uses the
+  top-level `lora` object. Style LoRAs like darkbrush need their trigger word in
+  the prompt.
+- Cache-DiT is supported; `has_separate_cfg` is checkpoint-aware (False for Turbo
+  no-CFG, True for Raw CFG).
+- Known limitations: TP / SP / CFG-Parallel are not yet wired for Krea 2. HSDP,
+  layerwise CPU offload, and VAE-patch-parallel (decode) are supported.
Author	SHA1	Message	Date
Gao Han	e8193787cc	Merge branch 'main' into fix/pr4709-review-followup	2026-07-18 11:28:29 +08:00
bastefaniak	d09f549e86	[bugfix] Use FlowMatchEulerDiscreteScheduler from diffusers 0.39.0 with correct RNG handling for upcoming Cosmos3 distilled models (#5176 ) Signed-off-by: Bartosz Stefaniak <bstefaniak@nvidia.com>	2026-07-17 11:38:04 -07:00
Weiming Liaoandwangyu	7aa5c9a090	[CI][NPU]: Add nightly performance test for NPU - Qwen3-TTS on A3 (#5158 ) Signed-off-by: Weiming Liao <liaowm5@gmail.com> Co-authored-by: wangyu <53896905+yenuo26@users.noreply.github.com>	2026-07-17 17:32:48 +08:00
linzhenpl07	be517db18a	examples: replica data-parallel recipe + benchmark for video DiT (#4707 ) (#5052 ) Signed-off-by: linzhenpl07 <linzhenpl07@gmail.com>	2026-07-17 03:26:35 +00:00
bastefaniakandRoger Wang	07957e12f4	Fix FP8 Quack GEMM call when inference_mode() and torch.compile() is used. (#5153 ) Signed-off-by: Bartosz Stefaniak <bstefaniak@nvidia.com> Co-authored-by: Roger Wang <hey@rogerw.io>	2026-07-16 15:01:21 -07:00
Alex Steiner	ae949d129f	[Bugfix][Quantization] Remap ModelOpt NVFP4 scale tensors (#5087 ) Signed-off-by: Alex Steiner <asteiner@nvidia.com>	2026-07-16 12:33:57 -07:00
bastefaniak	85f7a5b374	Add Cosmos3 Edge and Distilled checkpoints support (#5001 ) Signed-off-by: Bartosz Stefaniak <bstefaniak@nvidia.com>	2026-07-16 09:39:40 -07:00
zzh	0a9a5c231e	Fix hunyuan ci (#5143 ) Signed-off-by: zzh <943967662@qq.com>	2026-07-16 14:49:13 +00:00
wangyu	2a49b2b217	[CI][Perf] Add JSON mark support and paired parametrize for L4 perf, so as to further support performance‑case selection for NPU‑CI. (#5084 ) Signed-off-by: wangyu <410167048@qq.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com>	2026-07-16 14:45:00 +00:00
Canlin Guoandsunnyxyli	5c3900960b	[Feature] Support MOSS-TTS-Local-v1.5 vocoder graph (#4929 ) Signed-off-by: gcanlin <canlinguosdu@gmail.com> Signed-off-by: sunnyxyli <1550693033@qq.com> Co-authored-by: sunnyxyli <1550693033@qq.com>	2026-07-16 15:02:15 +08:00
Ricardo Noriega	68a992060b	[Bugfix] Fix audio.format parameter silently ignored in chat completions (#4718 ) Signed-off-by: Ricardo Noriega De Soto <rnoriega@redhat.com>	2026-07-16 00:17:02 -06:00
zhumingjue138	238fc0a609	[Bugfix] fix issue 5065 (#5118 )	2026-07-15 19:39:44 -07:00
Shi Boao Bowieandbjf-frz	a432b8a325	[Refactor][OutputProcessor 1/8]: clean up multimodal payload accumulation and move payload logic into MultimodalPayload & directory refactor (#4980 ) Signed-off-by: Boao Shi <aoibosh@connect.hku.hk> Signed-off-by: bjf-frz <frz123db@gmail.com> Co-authored-by: bjf-frz <frz123db@gmail.com>	2026-07-15 15:50:19 +08:00
wangyu	795e4e0160	[CI] Add local job runner scripts for L2-L4 pytest execution (#4672 ) Signed-off-by: wangyu <410167048@qq.com>	2026-07-15 14:18:58 +08:00
Honghan Zhu zhuhh97 Claude Opus 4.7 Gao Han Hongsheng Liu	f84f0e72fd	[Bug Fix] Hunyuan image cfg bugfix (#4752 ) Signed-off-by: zhuhh97 <zhuhonghan@huawei.com> Signed-off-by: dph97 <nenbaying@163.com> Co-authored-by: zhuhh97 <z00806815@gmail.com> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-07-15 13:08:15 +08:00
zyz111222andamy-why-3459	2d58070f60	[Hardware][Ascend][Model] Optimize Qwen3-TTS on 310P (#4841 ) Signed-off-by: zouyizhou <zouyizhou@huawei.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>	2026-07-15 08:57:44 +08:00
wkutak	a7b6a174bc	Fix loading of ModelOpt FP8 checkpoints for Cosmos3 (#5076 ) Signed-off-by: Wojciech Kutak <wkutak@nvidia.com>	2026-07-14 17:33:27 -07:00
wuli666andwuli666	6603450d71	[Bugfix][Quantization] Initialize component quantization base state (#5103 ) Signed-off-by: wuli666 <djjpro975@gmail.com> Co-authored-by: wuli666 <djjpro975@gmail.com>	2026-07-14 12:06:55 +00:00
bjf-frz	b14ff23041	[PERF][CI]Add Cosmos3 diffusion perf config (#5010 ) Signed-off-by: bjf-frz <frz123db@gmail.com>	2026-07-14 10:25:17 +00:00
bjf-frz	1a4a042a83	[REFACT]Refactor diffusion outputs to payload metadata (#4922 ) Signed-off-by: bjf-frz <frz123db@gmail.com>	2026-07-14 17:24:03 +08:00
Alex Steiner	4b9d09cb87	[Bugfix][Quantization] Support packed parameters with HSDP (#5088 ) Signed-off-by: Alex Steiner <asteiner@nvidia.com>	2026-07-14 08:24:26 +00:00
wjsuijlenh	3589a676e2	[Bugfix] Accept kv_prefetch_jobs in ARDiffusionModelRunner.execute_model (#4941 ) Signed-off-by: Wijnand Suijlen <wijnand.suijlen@huawei.com>	2026-07-14 08:22:40 +00:00
Zhang JianandChendi Xue	b1809985ef	[recipe] Add cosmos3-nano recipe for npu (1xA3) (#4978 ) Signed-off-by: Zhang Jian <jianmusings@gmail.com> Signed-off-by: Chendi Xue <chendi.xue@intel.com> Co-authored-by: Chendi Xue <chendi.xue@intel.com>	2026-07-14 16:19:44 +08:00
Diya Penganddiyapeng	e0570a0859	[Quantization] Add BitsAndBytes W4 online quantization for diffusion transformer (#5037 ) Signed-off-by: diyapeng <50388704+diyapeng@users.noreply.github.com> Co-authored-by: diyapeng <50388704+diyapeng@users.noreply.github.com>	2026-07-14 15:48:48 +08:00
Weiming Liao	b20ebc8f9b	[NPU] Upgrade CI IMAGE to v0.25.0 (#5095 ) Signed-off-by: Weiming Liao <liaowm5@gmail.com>	2026-07-14 15:47:02 +08:00
akshatvishu	2c0d8399ac	[Bugfix] Restore Qwen3-Omni speaker metadata during non-async handoff (#5086 ) Signed-off-by: akshatvishu <akshatnayak197@gmail.com>	2026-07-14 04:35:44 +00:00
Hongsheng Liuandhsliu_ustc	625892032b	[Doc] Standardize serving names and identifiers (#5081 ) Signed-off-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> Co-authored-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com>	2026-07-13 23:59:42 +08:00
Rahul Steiger	66b9ad287a	Extend Cosmos3 CPU offloading to component and layerwise paths (#4695 )	2026-07-13 09:48:53 +00:00
wangyu	98229b08ba	[Bugfix] Pass --trust-remote-code via HunyuanImage3 DFX serve_args (#5006 ) Signed-off-by: wangyu <410167048@qq.com>	2026-07-13 15:34:46 +08:00
Chendi.Xue	161b005cfc	[XPU][CI] Fix xpu after rebasing for 0.25.0 (#5066 ) Signed-off-by: Chendi Xue <chendi.xue@intel.com>	2026-07-13 15:34:27 +08:00
Zheng Wengang	3f4abedbda	[Perf] Ming-TTS Stage-0 flow-head speedup: RoPE/QKV fusion, opt-in torch.compile, PIECEWISE cudagraph (#4942 )	2026-07-12 19:50:51 -07:00
WeiQing Chen	8e14f69d96	docs: update WeChat QR code (#5061 ) Signed-off-by: David Chen <530634352@qq.com>	2026-07-13 10:24:17 +08:00
Yi Liu	9a02963acf	docs: document Cosmos3-Super AutoRound world-model support (#5062 ) Signed-off-by: yiliu30 <yi4.liu@intel.com>	2026-07-13 02:22:16 +00:00
Hongsheng Liu	a6da469134	docs: update README for v0.24.0 (#5060 ) Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-07-13 08:26:37 +08:00
LHXuuu	a28ded243f	[Perf][Ming-TTS] Add initial latent chunk for lower streaming TTFP (#5011 ) Signed-off-by: LHXuuu <xulianhao.xlh@antgroup.com>	2026-07-13 06:45:28 +08:00
Zhou TaichangandClaude Opus 4.8	d3c47efc4f	[Rebase] Rebase to vllm v0.25.0 (#5042 ) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>	2026-07-12 23:39:53 +08:00
wangyu	dda88a6f86	[BugFix] Align invalid layers DFX tests with 2-10 range (#5044 ) (#5047 ) Signed-off-by: wangyu <410167048@qq.com>	2026-07-12 09:50:54 +00:00
Alex Brooks	9014bb9917	[CI] Fix core_model Level for non-Diffusion Models (#5008 )	2026-07-12 03:00:50 +00:00
Naga Sai Abhinay Devarinti	6581e2afa9	[Model] Add Krea 2 text-to-image diffusion model (#4730 ) Signed-off-by: Abhinay1997 <devarintinagasaiabhinay@gmail.com>	2026-07-11 16:25:41 -07:00
tzhouamandClaude Fable 5	6d39fb25aa	[CI] Recalibrate rebase-raised timeouts on post-merge nightly durations Only post-#4709 nightlies (#2474/#2486, both on rebased code) are a valid baseline; the rebase changed step runtimes substantially. Revert the raises that post-merge data no longer supports and keep the ones it confirms: - Omni Function H100: keep 120m (post-merge 59-71 min vs old 90m ceiling) - Omni Perf No Async Chunk: 300m -> 180m (post-merge 41-44 min) - TTS Function L4: 120m -> 90m (post-merge ~27 min) - Diffusion X2I L4: 120m -> 60m (post-merge ~18 min) - IndexTTS2: 60m/45m -> 40m/35m (steady ~12.5 min warm, cold kill at 20.7m) - CustomVoice/Base 40m and Qwen3-Omni 50m: keep (9-21 min warm + cold-cache headroom the old limits lacked) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: tzhouam <tzhouam@connect.ust.hk>	2026-07-02 07:54:56 +00:00
tzhouamandClaude Fable 5	21f0986fd4	[CI] Ground timeout rationale comments in measured nightly durations Replace the generic timeout comments with actual job durations from the vllm-omni-rebase nightly pipeline (builds #2316-#2486, 2026-06-22..07-01): Omni H100 function test still timed out at the old 90m ceiling on 06-29; the no-async-chunk perf test passed at 179.4m against the old 180m limit the same night; TTS L4 passes ran 72-87m under the old 90m limit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: tzhouam <tzhouam@connect.ust.hk>	2026-07-02 07:49:45 +00:00
tzhouamandClaude Fable 5	35a423c9bd	[BugFix][CI] Address review follow-ups from v0.24.0 rebase (#4709 ) Addresses Gaohan123's post-approval review comments on #4709: - pipeline_dreamzero: restore platform-dispatched synchronize() instead of torch.cuda so DreamZero keeps working on non-CUDA hardware (regression of #4399 introduced during the rebase). - omni_scheduler_mixin: drop the VLLM_OMNI_INPUT_WAIT_TIMEOUT_S env-var override; the full-payload input-wait timeout is now a plain constant. - diffusion_worker: implement _DiffusionVllmModelConfig.is_diffusion by mirroring upstream ModelConfig.is_diffusion (dLLM canvas_length check) and document why omni DiT pipelines must not enable dLLM paths. - qwen3_tts: make _embedding_dtype a class attribute on the talker and prompt-embeds builder so __new__-based test stubs inherit it; remove the (duplicated) manual assignments from the preprocess test. - .buildkite: document why the rebase raised per-step timeouts (steps were killed by the old limits, masking real failures; e.g. CustomVoice passes in ~9-15 min but cold runs were killed at 20m). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: tzhouam <tzhouam@connect.ust.hk>	2026-07-02 07:36:14 +00:00