Fix LingBot diffusion alignment CI

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>

.buildkite/test-nightly.yml

@@ -650,8 +650,6 @@ 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:
@@ -695,7 +693,6 @@ 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:

.buildkite/test-ready.yml

@@ -684,24 +684,3 @@ steps:
                     hostPath:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
-
-      - label: "Diffusion · GGUF Plugin Test"
-        soft_fail: true
-        source_file_dependencies:
-          - tests/plugin/test_gguf_diffusion.py
-          - vllm_omni/diffusion/model_loader
-        commands:
-          - pip install "vllm-gguf-plugin>=0.0.3"
-          - timeout 30m pytest -sv tests/plugin/test_gguf_diffusion.py
-        agents:
-          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"

README.md

@@ -16,7 +16,6 @@ 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.
@@ -32,9 +31,9 @@ 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, audio, video, and action data processing
+- **Omni-modality**: Text, image, video, and audio 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 and action outputs
+- **Heterogeneous outputs**: from traditional text generation to multimodal outputs
 
 <p align="center">
   <picture>
@@ -58,10 +57,9 @@ 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, Cosmos3, HunyuanImage, BAGEL)
+- **Omni-modality models** (e.g. Qwen3-Omni, Cosmos, 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
 

benchmarks/diffusion/lingbot_video_parity.py

@@ -0,0 +1,217 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import statistics
+import subprocess
+import sys
+import time
+from pathlib import Path
+from typing import Any
+
+import imageio.v3 as iio
+import numpy as np
+import torch
+
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.platforms import current_omni_platform
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="LingBot dense parity and timing harness.")
+    parser.add_argument("--model", default="/home/models/lingbot-video-dense-1.3b")
+    parser.add_argument("--official-repo", default="/tmp/lingbot-video")
+    parser.add_argument("--output-dir", default="/tmp/lingbot_dense_parity")
+    parser.add_argument("--prompt", default="a robotic arm picks up a red block")
+    parser.add_argument("--height", type=int, default=192)
+    parser.add_argument("--width", type=int, default=320)
+    parser.add_argument("--num-frames", type=int, default=9)
+    parser.add_argument("--steps", type=int, default=2)
+    parser.add_argument("--guidance-scale", type=float, default=3.0)
+    parser.add_argument("--shift", type=float, default=3.0)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--fps", type=int, default=24)
+    parser.add_argument("--runs", type=int, default=0, help="Optional in-process steady-state native runs.")
+    return parser.parse_args()
+
+
+def _run(cmd: list[str], cwd: Path | None = None) -> tuple[float, str]:
+    start = time.perf_counter()
+    proc = subprocess.run(cmd, cwd=cwd, text=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+    elapsed = time.perf_counter() - start
+    if proc.returncode != 0:
+        print(proc.stdout)
+        raise subprocess.CalledProcessError(proc.returncode, proc.args, output=proc.stdout)
+    return elapsed, proc.stdout
+
+
+def _read_video(path: Path) -> np.ndarray:
+    frames = iio.imread(path, plugin="pyav")
+    if frames.dtype != np.float32:
+        frames = frames.astype(np.float32) / 255.0
+    return frames
+
+
+def _compare(a: np.ndarray, b: np.ndarray) -> dict[str, Any]:
+    if a.shape != b.shape:
+        return {"shape_a": list(a.shape), "shape_b": list(b.shape), "shape_match": False}
+    diff = a - b
+    mse = float(np.mean(diff * diff))
+    mae = float(np.mean(np.abs(diff)))
+    psnr = float("inf") if mse == 0 else float(20.0 * np.log10(1.0 / np.sqrt(mse)))
+    return {
+        "shape": list(a.shape),
+        "shape_match": True,
+        "mae": mae,
+        "mse": mse,
+        "psnr": psnr,
+    }
+
+
+def _extract_total_generation_seconds(log: str) -> float | None:
+    match = re.search(r"Total generation time:\s*([0-9.]+)\s*seconds", log)
+    return float(match.group(1)) if match else None
+
+
+def _peak_memory_mb(output: Any) -> float:
+    result = output[0] if isinstance(output, list) and output else output
+    return float(getattr(result, "peak_memory_mb", 0.0) or 0.0)
+
+
+def _benchmark_native_requests(args: argparse.Namespace) -> dict[str, Any]:
+    load_start = time.perf_counter()
+    omni = Omni(
+        model=args.model,
+        model_class_name="LingBotVideoPipeline",
+        flow_shift=args.shift,
+        parallel_config=DiffusionParallelConfig(),
+    )
+    load_seconds = time.perf_counter() - load_start
+    prompt = {"prompt": args.prompt}
+    latencies: list[float] = []
+    peak_mb = 0.0
+
+    for _ in range(args.runs):
+        generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(args.seed)
+        sampling_params = OmniDiffusionSamplingParams(
+            height=args.height,
+            width=args.width,
+            num_frames=args.num_frames,
+            num_inference_steps=args.steps,
+            guidance_scale=args.guidance_scale,
+            generator=generator,
+            output_type="pt",
+            extra_args={"shift": args.shift},
+        )
+        start = time.perf_counter()
+        output = omni.generate(prompt, sampling_params)
+        latencies.append(time.perf_counter() - start)
+        peak_mb = max(peak_mb, _peak_memory_mb(output))
+        del output
+    omni.close()
+
+    return {
+        "load_seconds": load_seconds,
+        "run_seconds": latencies,
+        "mean_seconds": statistics.mean(latencies) if latencies else None,
+        "median_seconds": statistics.median(latencies) if latencies else None,
+        "min_seconds": min(latencies) if latencies else None,
+        "max_seconds": max(latencies) if latencies else None,
+        "peak_mb": peak_mb,
+    }
+
+
+def main() -> None:
+    args = parse_args()
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    official_out = output_dir / "official_diffusers.mp4"
+    native_out = output_dir / "vllm_omni_native.mp4"
+
+    official_cmd = [
+        sys.executable,
+        str(Path(args.official_repo) / "scripts" / "inference.py"),
+        "--backend",
+        "diffusers",
+        "--model_dir",
+        args.model,
+        "--mode",
+        "t2v",
+        "--prompt",
+        args.prompt,
+        "--output",
+        str(official_out),
+        "--height",
+        str(args.height),
+        "--width",
+        str(args.width),
+        "--num_frames",
+        str(args.num_frames),
+        "--steps",
+        str(args.steps),
+        "--guidance_scale",
+        str(args.guidance_scale),
+        "--shift",
+        str(args.shift),
+        "--seed",
+        str(args.seed),
+        "--fps",
+        str(args.fps),
+        "--quiet_progress",
+    ]
+    native_cmd = [
+        sys.executable,
+        "examples/offline_inference/text_to_video/text_to_video_lingbot.py",
+        "--model",
+        args.model,
+        "--prompt",
+        args.prompt,
+        "--output",
+        str(native_out),
+        "--height",
+        str(args.height),
+        "--width",
+        str(args.width),
+        "--num-frames",
+        str(args.num_frames),
+        "--num-inference-steps",
+        str(args.steps),
+        "--guidance-scale",
+        str(args.guidance_scale),
+        "--flow-shift",
+        str(args.shift),
+        "--seed",
+        str(args.seed),
+        "--fps",
+        str(args.fps),
+    ]
+
+    official_time, official_log = _run(official_cmd, cwd=Path(args.official_repo))
+    native_time, native_log = _run(native_cmd, cwd=Path.cwd())
+
+    result = {
+        "official_seconds": official_time,
+        "native_process_seconds": native_time,
+        "native_request_seconds": _extract_total_generation_seconds(native_log),
+        "official_output": str(official_out),
+        "native_output": str(native_out),
+        "accuracy": _compare(_read_video(official_out), _read_video(native_out)),
+        "official_log_tail": official_log.splitlines()[-20:],
+        "native_log_tail": native_log.splitlines()[-20:],
+    }
+    if args.runs > 0:
+        result["native_steady_state"] = _benchmark_native_requests(args)
+
+    result_path = output_dir / "comparison.json"
+    result_path.write_text(json.dumps(result, indent=2, sort_keys=True) + "\n")
+    print(json.dumps(result, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()

docker/Dockerfile.ci

@@ -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=v0.25.0
+ARG VLLM_BASE_TAG=v0.24.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}

docs/README.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, audio, video, and action data processing
+- **Omni-modality**: Text, image, video, and audio 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 and action outputs
+- **Heterogeneous outputs**: from traditional text generation to multimodal outputs
 
 <p align="center">
   <picture>
@@ -55,10 +55,8 @@ 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, Cosmos3, HunyuanImage, BAGEL)
+- Omni-modality models (e.g. Qwen2.5-Omni, Qwen3-Omni)
-- TTS models (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
+- Multi-modality generation models (e.g. Qwen-Image)
-- 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:
 

docs/configuration/composable_parallel.md

@@ -1,239 +0,0 @@
-# Composable parallel strategies
-
-In vLLM-Omni, the `composable_parallel` system layers a declarative *parallel strategy* on top of the existing stage configs. A strategy YAML maps each pipeline stage (by its `model_stage` name) to a stack of axis declarations — one per parallelism dimension — that the engine translates into per-stage parallel sizing (`tensor_parallel_size`, `data_parallel_size`, `pipeline_parallel_size`, expert-parallel toggle, replica count) and, when relevant, an orchestrator-wide load-balancer policy.
-
-Strategies are intentionally narrow: each one only writes the axes it declares, leaving every other engine arg untouched. This makes a strategy file a small, reusable overlay you can swap independently of the deploy YAML (see `configuration/stage_configs.md`).
-
-!!! note
-    Composable parallel is **opt-in**. Passing `--strategy-config <path/to/strategy.yaml>` overlays the derived sizing onto the registry-merged stages *before* any CLI overrides are applied. When `--strategy-config` is omitted, stages remain exactly as their deploy YAML produces them.
-
-## Top-level schema reference
-
-A strategy file has a single top-level mapping, `strategies`, keyed by the `model_stage` name of each stage you want to overlay. Each value is a list of axis declarations (see [StrategySpec fields](#strategyspec-fields) below).
-
-| Field | Type | Required | Default | Description |
-|-------|------|----------|---------|-------------|
-| `strategies` | `dict[str, list]` | required | — | Maps `model_stage` name (e.g. `thinker`, `talker`, `code2wav`, `dit`) to a list of axis declarations. The key MUST be the stage name, not the integer `stage_id`. See `vllm_omni/config/composable_parallel/strategy_loader.py`. |
-
-### StrategySpec fields
-
-Each entry in a stage's list becomes one `StrategySpec` (see `vllm_omni/config/composable_parallel/spec.py`). The loader fills in `routing` and `aggregation` from per-kind defaults, so the file stays compact.
-
-| Field | Type | Required | Default | Description |
-|-------|------|----------|---------|-------------|
-| `axis` | str | required | — | Mesh-axis kind. One of `tp`, `dp`, `pp`, `ep`, `stage_replica` (wired today) or `sp_ulysses`, `sp_ring`, `cfg`, `vae_pp`, `hsdp`, `stage_pp`, `cp` (reserved; declaring one raises `AxisTranslationError` at translate time). See `MeshAxisKind` in `vllm_omni/config/composable_parallel/spec.py`. |
-| `size` | int | required | — | Axis degree. Must be `> 0`. |
-| `routing` | str | optional | per-kind default | Stateless routing policy. Accepted only for `dp` / `stage_replica`; other kinds reject it. Values: `random`, `round_robin`, `least_queue`. (`hash` is recognised but currently raises `NotImplementedError`.) See `RouteByStage` in `vllm_omni/config/composable_parallel/routing.py`. |
-| `l1_owner` | str | optional | per-kind default | Overrides which layer owns request routing for this axis. `engine` for `tp`/`dp`/`pp`/`ep`, `delegated` for `stage_replica`. Recorded into `shard_extension`. |
-| `name` | str | optional | axis kind | Human-readable label for this spec, used in error messages. |
-
-Two additional `StrategySpec` slots — `layer_hook_specs` and `kernel_specs` — let advanced authors bind layer-walker hook IDs (`LayerHookSpec`) or kernel IDs (`KernelSpec`) onto an axis. These are not exposed by the YAML loader today and are intended for programmatic construction. See `vllm_omni/config/composable_parallel/spec.py` for the dataclasses.
-
-### CLI flags
-
-| Flag | Description |
-|------|-------------|
-| `--strategy-config PATH` | Path to a `strategy.yaml`. Only takes effect on registry-based deploy paths (`--deploy-config` or the bundled default deploy YAML). Its derived sizing is overlaid onto the merged stages *before* CLI overrides. |
-| `--stage-configs-path PATH` | **Legacy.** Loads the old `stage_args` YAML schema for un-migrated models. Passing `--strategy-config` together with this path is silently inapplicable — vLLM-Omni emits a warning and the strategy is ignored (see `vllm_omni/entrypoints/utils.py`, lines 372–380). |
-| `--omni-lb-policy POLICY` | Orchestrator-wide load-balancer policy. Strategy-derived from any `stage_replica` axis; an explicit CLI value wins and must match the derived one or `AsyncOmniEngine` raises `Conflicting load-balancer policy` at construction. |
-
-!!! warning
-    `--strategy-config` cannot be used with `--stage-configs-path` (the legacy deploy path). If your model has not been migrated to the registry-based pipeline yet, the strategy is silently dropped with a warning. Migrate to a registry-based model or use `--deploy-config` to apply a strategy.
-
-## Sub-schemas
-
-### `axis`
-
-The set of recognised axis kinds is the `MeshAxisKind` enum in `vllm_omni/config/composable_parallel/spec.py`:
-
-```yaml
-# Translatable today (see `_SUPPORTED_KINDS` in translator.py).
-axis: tp              # tensor parallel — sets tensor_parallel_size
-axis: dp              # engine data parallel — sets data_parallel_size (vLLM's intra-engine DP LB)
-axis: pp              # pipeline parallel — sets pipeline_parallel_size
-axis: ep              # dense expert parallel — sets enable_expert_parallel=True; size must equal tp*dp
-axis: stage_replica   # omni-coordinator-level replica fan-out — sets num_replicas + omni_lb_policy
-
-# Reserved (raise AxisTranslationError today; declared in the type system for forward compat).
-axis: sp_ulysses
-axis: sp_ring
-axis: cfg
-axis: vae_pp
-axis: hsdp
-axis: stage_pp
-axis: cp
-```
-
-### `routing`
-
-The full closed hierarchy lives in `vllm_omni/config/composable_parallel/routing.py` (`Broadcast`, `PartitionByHash`, `ShardSequence`, `ShardSpatial`, `RouteByExpert`, `PipelineMicrobatch`, `DuplicateWithCondTag`, `RouteByStage`). For YAML authors only `RouteByStage` is exposed as a `routing:` string, and only for `dp` / `stage_replica` axes (other axes have fixed routing — `tp`/`ep` broadcast, `pp` microbatch). Recognised policies:
-
-```yaml
-routing: random         # uniform-random across replicas/DP ranks
-routing: round_robin    # cycle through replicas/DP ranks (the default for dp/stage_replica)
-routing: least_queue    # send to the replica with the smallest backlog
-# routing: hash         # recognised but raises NotImplementedError; no key-stable balancer yet
-```
-
-### `aggregation`
-
-Aggregation is filled in automatically by the loader (`_default_aggregation` in `vllm_omni/config/composable_parallel/strategy_loader.py`) from the axis kind, so users do not author it:
-
-| Axis kind | Aggregation pattern |
-|-----------|---------------------|
-| `tp`, `ep` | `TakeRank` (all ranks agree; pick one) |
-| `dp`, ep (in some flows) | `Union` (disjoint replica results) |
-| `pp` | `StitchPipeline` |
-| `stage_replica` | `FanInByStage` |
-
-Full pattern hierarchy: `TakeRank`, `Union`, `GatherDim`, `AllGather`, `StitchSpatial`, `StitchPipeline`, `Combine`, `FanInByStage` (see `vllm_omni/config/composable_parallel/aggregation.py`).
-
-## Precedence
-
-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`).
-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.
-
-## Worked example
-
-### Example 1 — `strategy_tp2.yaml`
-
-Shard the `thinker` stage with TP=2 on top of the bundled Qwen2.5-Omni deploy YAML. Routing is fixed (broadcast) for TP, so no `routing:` field is needed.
-
-```yaml
-# examples/offline_inference/qwen2_5_omni/strategy_tp2.yaml
-strategies:
-  thinker:
-    - axis: tp
-      size: 2
-```
-
-Launch:
-
-```bash
-python3 end2end.py --query-type text --num-prompts 6 \
-    --strategy-config strategy_tp2.yaml \
-    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
-```
-
-Effective `OmniParallelConfig` per stage after the overlay (see `OmniParallelConfig` in `vllm_omni/config/composable_parallel/translator.py`):
-
-| Stage (`model_stage`) | `tensor_parallel_size` | `data_parallel_size` | `pipeline_parallel_size` | `stage_replica_size` | `omni_lb_policy` |
-|-----------------------|------------------------|----------------------|--------------------------|----------------------|------------------|
-| `thinker`             | **2** (from strategy)  | 1                    | 1                        | 1                    | —                |
-| `talker`              | 1                      | 1                    | 1                        | 1                    | —                |
-| `code2wav`            | 1                      | 1                    | 1                        | 1                    | —                |
-
-### Example 2 — `strategy_stage_replica.yaml`
-
-Replicate the `thinker` stage across two independent engines and let omni's StagePool round-robin requests across them.
-
-```yaml
-# examples/offline_inference/qwen2_5_omni/strategy_stage_replica.yaml
-strategies:
-  thinker:
-    - axis: stage_replica
-      size: 2
-      routing: round_robin
-```
-
-Launch:
-
-```bash
-python3 end2end.py --query-type text --num-prompts 6 \
-    --strategy-config strategy_stage_replica.yaml \
-    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
-```
-
-Effective per-stage config and the derived orchestrator knob:
-
-| Stage (`model_stage`) | `tensor_parallel_size` | `stage_replica_size` | `omni_lb_policy` (derived) |
-|-----------------------|------------------------|----------------------|----------------------------|
-| `thinker`             | 1                      | **2**                | **`round-robin`**          |
-| `talker`              | 1                      | 1                    | —                          |
-| `code2wav`            | 1                      | 1                    | —                          |
-
-The derived `omni_lb_policy = round-robin` is **not** written into any stage's engine args. Instead it is surfaced on `StrategyApplyResult.omni_lb_policy` and applied once at orchestrator construction by `AsyncOmniEngine._apply_strategy_lb_policy` (see `vllm_omni/engine/async_omni_engine.py`, lines 1107–1134). If you also pass `--omni-lb-policy round-robin` the values must match; passing a different value raises `Conflicting load-balancer policy`.
-
-## Per-field deep-dive
-
-### `name`
-
-Default: the axis kind (e.g. `"tp"`).
-
-Human-readable label that appears in `AxisTranslationError` and `StrategyApplyError` messages. Set it to disambiguate two stages declaring the same axis kind when scanning logs.
-
-### `mesh_axis` (`axis` + `size`)
-
-Default: no default — both fields are required.
-
-`axis` chooses the parallelism dimension; `size` is its degree. Five kinds are wired end-to-end today (`tp`, `dp`, `pp`, `ep`, `stage_replica`); the remaining kinds in `MeshAxisKind` are reserved and raise `AxisTranslationError` at translate time so a typo cannot silently become a no-op (see `_SUPPORTED_KINDS` in `vllm_omni/config/composable_parallel/translator.py`). For `ep`, the size MUST equal `tensor_parallel_size * data_parallel_size`.
-
-### `routing`
-
-Default: per-kind — `RouteByStage(round_robin)` for `dp` and `stage_replica`; `PipelineMicrobatch()` for `pp`; `Broadcast()` for `tp`/`ep` and any other kind.
-
-User-facing string only on `dp` and `stage_replica` axes. The four recognised values are `random`, `round_robin`, `least_queue`, `hash`; `hash` is rejected today with `NotImplementedError`. See `RouteByStage` in `vllm_omni/config/composable_parallel/routing.py`.
-
-### `l1_owner`
-
-Default: per-kind — `engine` for `tp`/`dp`/`pp`/`ep`, `delegated` for `stage_replica`.
-
-Overrides which layer owns request routing along this axis. `engine` means vLLM realizes routing intra-engine; `delegated` means the omni coordinator owns routing across independent replicas. Set this only when overriding the per-kind default in `_DEFAULT_L1_OWNER` (`vllm_omni/config/composable_parallel/translator.py`); invalid values fail fast in `_resolve_l1_owner`.
-
-### `stage_replica` and `omni_lb_policy`
-
-`stage_replica` is *not* a vLLM world dimension; it spins up `size` independent engine replicas of one stage, coordinated by the omni `StagePool` load balancer. The axis writes both:
-
-- The stage's `num_replicas` (a per-stage deploy knob).
-- The pipeline-wide `omni_lb_policy` string (an orchestrator-level knob).
-
-Because `omni_lb_policy` is read once at `AsyncOmniEngine` construction — not from per-stage engine args — only one value is allowed across the whole pipeline. If two stages declare conflicting policies, application raises (see `apply_strategy_specs` in `vllm_omni/config/composable_parallel/apply.py`, lines 253–260). To set this knob manually instead, use `--omni-lb-policy` on the CLI; see [the model stage docs](stage_configs.md) for how stage replicas plug into the orchestrator at runtime.
-
-### `model_stage` keys
-
-Default: no default — required for every entry under `strategies:`.
-
-Strategy keys MUST be `model_stage` *names* (strings), the same labels stage configs use for `engine_args.model_stage` (see [`model_stage` in `stage_configs.md`](stage_configs.md#engine_argsmodel_stage)). Integer `stage_id` values are not accepted; the resolver matches on `model_stage` exactly (`_resolve_stage` in `vllm_omni/config/composable_parallel/apply.py`).
-
-## Common pitfalls
-
-!!! warning "Using `stage_id` integers as keys"
-    Strategy keys are `model_stage` names (str), not integer stage IDs. A file that uses an integer index will raise at apply time:
-
-    ```text
-    StrategyApplyError: strategy role model_stage=0 did not match any stage;
-    available stages: 'thinker'(id=0), 'talker'(id=1), 'code2wav'(id=2)
-    ```
-
-    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.
-
-!!! warning "Passing `--strategy-config` against the legacy `--stage-configs-path`"
-    Composable parallel only applies on the registry-based deploy path. If a model still resolves through the legacy `stage_args` YAML (loaded via `--stage-configs-path`), `--strategy-config` is silently inapplicable and emits:
-
-    ```text
-    --strategy-config (<path>) was provided but model '<model>' resolves via the
-    legacy stage_configs YAML path, which does not support composable-parallel
-    strategies; the strategy is ignored. Use a registry-based model to apply it.
-    ```
-
-    Migrate the model to the new deploy schema (see `configuration/stage_configs.md`) before applying a strategy.
-
-!!! warning "Conflicting `omni_lb_policy` across stages"
-    `omni_lb_policy` is a single pipeline-wide knob. If two `stage_replica` axes in different stages derive different policies, `apply_strategy_specs` raises:
-
-    ```text
-    StrategyApplyError: role 'talker' derives omni_lb_policy='least-queue-length'
-    but role 'thinker' already derived 'round-robin'. The load-balancer policy is
-    pipeline-wide; only one value is allowed.
-    ```
-
-    Make the policies match, drop one, or supply the value explicitly via `--omni-lb-policy`.

docs/configuration/pd_disaggregation.md

@@ -123,7 +123,7 @@ After inserting the extra thinker stage, renumber the remaining stages:
     runtime:
       devices: "2"
     engine_input_source: [1]
-    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_token_only
+    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker
 
   - stage_id: 3
     runtime:
@@ -131,7 +131,7 @@ After inserting the extra thinker stage, renumber the remaining stages:
     engine_args:
       max_num_seqs: 1
     engine_input_source: [2]
-    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav_token_only
+    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav
 ```
 
 Compared with the default Qwen3-Omni config:

docs/contributing/model/adding_omni_model.md

@@ -475,64 +475,92 @@ thinker_hidden_states = output.multimodal_output["24"]
 
 ### Implementation Example
 
-Create stage transition processors in `vllm_omni/model_executor/stage_input_processors/your_model_name.py`. Each inter-stage edge should provide a **coherent processor set** rather than a single monolithic function:
+Create stage transition processors in `vllm_omni/model_executor/stage_input_processors/your_model_name.py`:
-
-| Suffix | Role | Runs when |
-|--------|------|-----------|
-| `*_full_payload` | Worker-side payload producer | `async_chunk=false`; accumulates tensors and ships via connector |
-| `*_async_chunk` | Scheduler-side streaming producer | `async_chunk=true`; emits per-chunk payloads |
-| `*_token_only` | Orchestrator placeholder builder | `async_chunk=false`; allocates downstream prompt slots only |
 
 ```python
-# qwen3_omni.py (Thinker → Talker, non-async path)
+# qwen3_omni.py
 
-def thinker2talker_token_only(
+def thinker2talker(
-    source_outputs: list[Any],
+    stage_list: list[Any],
+    engine_input_source: list[int],
     prompt: OmniTokensPrompt | TextPrompt | None = None,
     requires_multimodal_data: bool = False,
-    streaming_context: Any | None = None,
 ) -> list[OmniTokensPrompt]:
-    """Allocate talker prefill slots; bulk tensors arrive via the connector."""
+    """
-    ...
+    Process thinker outputs to create talker inputs.
+
+    Args:
+        stage_list: List of stage objects
+        engine_input_source: Source stage IDs (typically [0] for thinker)
+        prompt: Original prompt data
+
+    Returns:
+        List of OmniTokensPrompt for talker stage
+    """
+    source_stage_id = engine_input_source[0]
+    thinker_outputs = stage_list[source_stage_id].engine_outputs
+    talker_inputs = []
+
+    for thinker_output in thinker_outputs:
+        output = thinker_output.outputs[0]
+        # Extract thinker embeddings and hidden states
+        thinker_prefill_embeddings = output.multimodal_output["0"].float().clone().detach().cuda()
+        thinker_hidden_states = output.multimodal_output["24"].float().clone().detach().cuda()
+
+        info = {
+            "thinker_prefill_embeddings": thinker_prefill_embeddings,
+            "thinker_hidden_states": thinker_hidden_states,
+            "thinker_sequences": thinker_output.prompt_token_ids + output.token_ids,
+            "thinker_input_ids": thinker_output.prompt_token_ids,
+        }
+
+        talker_inputs.append(
+            OmniTokensPrompt(
+                prompt_token_ids=[0] * computed_length,
+                additional_information=info,
+                multi_modal_data=None,
+            )
+        )
+
+    return talker_inputs
 
 
-def thinker2talker_full_payload(
+def talker2code2wav(
-    transfer_manager: Any,
+    stage_list: list[Any],
-    pooling_output: dict[str, Any],
+    engine_input_source: list[int],
-    request: OmniEngineCoreRequest,
+    prompt: OmniTokensPrompt | TextPrompt | None = None,
-) -> dict[str, Any] | None:
+    requires_multimodal_data: bool = False,
-    """Pack accumulated thinker hidden states into OmniPayload for stage-1."""
+) -> list[OmniTokensPrompt]:
-    ...
+    """
+    Process talker outputs to create code2wav inputs.
+    """
+    source_stage_id = engine_input_source[0]
+    talker_outputs = stage_list[source_stage_id].engine_outputs
+    code2wav_inputs = []
 
+    for talker_output in talker_outputs:
+        output = talker_output.outputs[0]
+        # Extract codec codes
+        codec_codes = (
+            output.multimodal_output["code_predictor_codes"]
+            .to(torch.long)
+            .transpose(0, 1)
+            .cpu()
+            .to(torch.long)
+            .reshape(-1)
+            .tolist()
+        )
 
-def thinker2talker_async_chunk(
+        code2wav_inputs.append(
-    transfer_manager: Any,
+            OmniTokensPrompt(
-    multimodal_output: OmniPayload | dict[str, Any],
+                prompt_token_ids=codec_codes,
-    request: OmniEngineCoreRequest,
+                multi_modal_data=None,
-    is_finished: bool = False,
+            )
-) -> OmniPayloadStruct | None:
+        )
-    """Stream thinker rows to talker while async_chunk is enabled."""
+
-    ...
+    return code2wav_inputs
 ```
 
-Wire these in `pipeline.py`:
-
-```python
-StagePipelineConfig(
-    stage_id=0,
-    custom_process_next_stage_input_func=f"{_PROC}.thinker2talker_full_payload",
-    async_chunk_process_next_stage_input_func=f"{_PROC}.thinker2talker_async_chunk",
-    ...
-),
-StagePipelineConfig(
-    stage_id=1,
-    sync_process_input_func=f"{_PROC}.thinker2talker_token_only",
-    ...
-),
-```
-
-Do **not** add a no-suffix `thinker2talker` when a `sync_process_input_func` is already declared — `_select_processor_funcs()` always prefers the `*_token_only` hook in non-async mode, so the bare function would never run. See `docs/design/rfc_stage_input_processors_refactor.md` for the full contract.
-
 ## Testing
 
 For comprehensive testing guidelines, please refer to the [Test File Structure and Style Guide](../ci/tests_style.md).

docs/design/architecture_overview.md

@@ -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, video, and action trajectories, alongside text.
+* **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-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.

docs/getting_started/installation/gpu.md

@@ -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, 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).
+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).
 
 === "NVIDIA CUDA"
 

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

@@ -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==0.25.0 --torch-backend=auto
+uv pip install vllm==0.24.0 --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 0.25.0 release version of the library
+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
 
 ```bash
-uv pip install vllm==0.25.0 --torch-backend=auto
+uv pip install vllm==0.24.0 --torch-backend=auto
 ```
 
-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.
+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.
 
 #### 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 v0.25.0
+git checkout v0.24.0
 ```
 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 v0.25.0; the wheel filename has no `+cu130` suffix)
+#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/v0.25.0/vllm-0.25.0-cp38-abi3-manylinux_2_28_x86_64.whl
+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
 ```
 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:v0.25.0 \
+    vllm/vllm-omni:v0.24.0 \
     vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 

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

@@ -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==0.25.0+rocm723 --extra-index-url https://wheels.vllm.ai/rocm/0.25.0/rocm723
+uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
 ```
 
 #### 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 0.25.0 release version of the library
+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
 
 ```bash
-uv pip install vllm==0.25.0+rocm723 --extra-index-url https://wheels.vllm.ai/rocm/0.25.0/rocm723
+uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
 ```
 
-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.
+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.
 
 #### 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 v0.25.0
+git checkout v0.24.0
 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:v0.25.0 \
+  vllm/vllm-omni-rocm:v0.24.0 \
   --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:v0.25.0
+  vllm/vllm-omni-rocm:v0.24.0
 ```
 
 # --8<-- [end:pre-built-images]

docs/getting_started/quickstart.md

@@ -19,10 +19,10 @@ uv venv --python 3.12 --seed
 source .venv/bin/activate
 
 # On CUDA
-uv pip install vllm==0.25.0 --torch-backend=auto
+uv pip install vllm==0.24.0 --torch-backend=auto
 
 # On ROCm
-uv pip install vllm==0.25.0+rocm723 --extra-index-url https://wheels.vllm.ai/rocm/0.25.0/rocm723
+uv pip install vllm==0.24.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.24.0/rocm722
 
 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 < `0.25.0` and vLLM Omni `0.25.0`, 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 < `0.24.0` and vLLM Omni `0.24.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
 
 ## Offline Inference
 

docs/models/supported_models.md

@@ -34,7 +34,6 @@ 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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
@@ -47,6 +46,7 @@ th {
 | `LTX2ImageToVideoTwoStagesPipeline` | LTX-2-I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | |
 | `LTX23Pipeline` | LTX-2.3-T2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
 | `LTX23ImageToVideoPipeline` | LTX-2.3-I2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
+| `LingBotVideoPipeline` | LingBot-Video dense T2V | `robbyant/lingbot-video-dense-1.3b` | ✅︎ | | | |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | |
 | `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | |

docs/user_guide/diffusion_features.md

@@ -121,7 +121,6 @@ The following tables show which models support each feature:
 | **GLM-Image**            |     ❌     |     ❌      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Hidream-I1-Full**        |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **HunyuanImage3**        |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ✅        |        ❌         |
-| **Krea 2**               |     ❌     |     ✅      |           ❌           |       ❌        |         ❌         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LongCat-Image**        |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **LongCat-Image-Edit**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **MagiHuman**            |     ❌     |     ❌      |           ❌           |       ❓        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -145,7 +144,6 @@ The following tables show which models support each feature:
 > 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.
-> 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
 

docs/user_guide/quantization/gguf.md

@@ -9,15 +9,6 @@ for the transformer.
 
 GGUF is static quantization: the quantized weights are produced before serving.
 
-!!! note
-    GGUF support has migrated to OOT [vllm-gguf-plugin](https://github.com/vllm-project/vllm-gguf-plugin). Make sure you have GGUF plugin installed before serving a GGUF model.
-
-Before serving a GGUF model, make sure to install the [vllm-gguf-plugin](https://github.com/vllm-project/vllm-gguf-plugin):
-
-```bash
-uv pip install vllm-gguf-plugin>=0.0.3
-```
-
 ## Hardware Support
 
 | Device | Support |

examples/offline_inference/image_to_video/image_to_video.py

@@ -220,7 +220,7 @@ def parse_args() -> argparse.Namespace:
         "--quantization",
         type=str,
         default=None,
-        choices=["fp8", "mxfp8", "mxfp4", "mxfp4_dualscale", "int8"],
+        choices=["fp8", "mxfp8", "mxfp4", "mxfp4_dualscale", "int8", "gguf"],
         help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 (NPU). mxfp4: W4A4 MXFP4 (NPU). mxfp4_dualscale: W4A4 MXFP4 dual-scale + BF16 fallback mixed (NPU). fp8: online FP8 (GPU).",
     )
 

examples/offline_inference/qwen2_5_omni/README.md

@@ -60,24 +60,3 @@ If media file paths are not provided, the script will use default assets. Suppor
 - `mixed_modalities`: Audio + image + video
 - `use_audio_in_video`: Extract audio from video
 - `text`: Text-only query
-
-### Composable parallelism (strategy configs)
-
-You can shard or replicate a stage with a small composable-parallel
-`strategy.yaml` overlaid onto the **bundled default deploy config** -- no bespoke
-deploy YAML required. Supply the strategy with `--strategy-config` and any matching
-device layout via `--stage-overrides` (explicit kwargs forwarded to `Omni`).
-
-Both examples below require >= 3 GPUs (the thinker uses 2, talker + code2wav share a third):
-
-```bash
-# Tensor-parallel the thinker (TP=2)
-python end2end.py --query-type text --num-prompts 6 \
-    --strategy-config strategy_tp2.yaml \
-    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
-
-# Replicate the thinker across 2 engines (round-robin load balancing)
-python end2end.py --query-type text --num-prompts 6 \
-    --strategy-config strategy_stage_replica.yaml \
-    --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
-```

examples/offline_inference/qwen2_5_omni/end2end.py

@@ -553,39 +553,6 @@ def parse_args():
         default=False,
         help="Use py_generator mode. The returned type of Omni.generate() is a Python Generator object.",
     )
-    parser.add_argument(
-        "--deploy-config",
-        type=str,
-        default=None,
-        help="Optional explicit deploy YAML (otherwise the bundled default is used).",
-    )
-    parser.add_argument(
-        "--strategy-config",
-        type=str,
-        default=None,
-        help=(
-            "Optional composable-parallel strategy.yaml mapping role -> parallel axes "
-            "(e.g. tp / stage_replica). Overlaid onto the merged stage configs."
-        ),
-    )
-    parser.add_argument(
-        "--stage-overrides",
-        type=str,
-        default=None,
-        help=(
-            "Optional JSON of per-stage overrides applied on top of the default "
-            'deploy config, e.g. \'{"0": {"devices": "0,1"}}\' to give the '
-            "thinker a 2-GPU pool. Lets you run a strategy on the bundled default "
-            "deploy config without writing a bespoke deploy YAML."
-        ),
-    )
-    parser.add_argument(
-        "--omni-lb-policy",
-        type=str,
-        default=None,
-        choices=["random", "round-robin", "least-queue-length"],
-        help="StagePool load-balancer policy for replicated stages (orchestrator-level knob).",
-    )
     return parser.parse_args()
 
 

examples/offline_inference/qwen2_5_omni/strategy_stage_replica.yaml

@@ -1,22 +0,0 @@
-# Composable-parallel strategy for Qwen2.5-Omni.
-#
-# Replicate the thinker stage across 2 independent engines and let the omni
-# StagePool balance requests round-robin across them. The thinker's `devices`
-# entry acts as a per-replica template, so omni expands it to one GPU per replica.
-# The `routing: round_robin` below auto-derives the StagePool load-balancer policy,
-# so --omni-lb-policy is not needed.
-#
-# Rather than hand-write a bespoke deploy YAML, run on the bundled default deploy
-# config and pass the matching 2-GPU device template via --stage-overrides (explicit
-# kwargs): the thinker gets a two-GPU pool while talker + code2wav share a third
-# card. Requires >= 3 GPUs.
-#
-# Usage:
-#   python3 end2end.py --query-type text --num-prompts 6 \
-#       --strategy-config strategy_stage_replica.yaml \
-#       --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
-strategies:
-  thinker:
-    - axis: stage_replica
-      size: 2
-      routing: round_robin

examples/offline_inference/qwen2_5_omni/strategy_tp2.yaml

@@ -1,19 +0,0 @@
-# Composable-parallel strategy for Qwen2.5-Omni: tensor-parallel the thinker.
-#
-# Shard the thinker stage across 2 GPUs with tensor parallelism (TP=2). TP routing
-# is fixed (broadcast), so no `routing` / load-balancer policy is involved -- this
-# is the simplest axis and a good starting point for the mental model.
-#
-# The strategy sets tensor_parallel_size=2 on the thinker. Rather than hand-write a
-# bespoke deploy YAML, run on the bundled default deploy config and pass a matching
-# 2-GPU device layout via --stage-overrides (explicit kwargs), so the thinker gets a
-# two-GPU pool while talker + code2wav share a third card. Requires >= 3 GPUs.
-#
-# Usage:
-#   python3 end2end.py --query-type text --num-prompts 6 \
-#       --strategy-config strategy_tp2.yaml \
-#       --stage-overrides '{"0": {"devices": "0,1"}, "1": {"devices": "2"}, "2": {"devices": "2"}}'
-strategies:
-  thinker:
-    - axis: tp
-      size: 2

examples/offline_inference/text_to_image/README.md

@@ -36,7 +36,6 @@ 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.
@@ -197,37 +196,6 @@ 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`

examples/offline_inference/text_to_image/text_to_image.py

@@ -202,11 +202,17 @@ def parse_args() -> argparse.Namespace:
         "--quantization",
         type=str,
         default=None,
-        choices=["fp8", "int8"],
+        choices=["fp8", "int8", "gguf"],
         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), 'gguf' (GGUF quantized weights). "
         "Default: None (no quantization, uses BF16).",
     )
+    parser.add_argument(
+        "--gguf-model",
+        type=str,
+        default=None,
+        help=("GGUF file path or HF reference for transformer weights. Required when --quantization gguf is set."),
+    )
     parser.add_argument(
         "--ignored-layers",
         type=str,
@@ -396,7 +402,14 @@ def main():
     # ignored_layers is specified so the list flows through OmniDiffusionConfig
     quant_kwargs: dict[str, Any] = {}
     ignored_layers = [s.strip() for s in args.ignored_layers.split(",") if s.strip()] if args.ignored_layers else None
-    if args.quantization and ignored_layers:
+    if args.quantization == "gguf":
+        if not args.gguf_model:
+            raise ValueError("--gguf-model is required when --quantization gguf is set.")
+        quant_kwargs["quantization_config"] = {
+            "method": "gguf",
+            "gguf_model": args.gguf_model,
+        }
+    elif args.quantization and ignored_layers:
         quant_kwargs["quantization_config"] = {
             "method": args.quantization,
             "ignored_layers": ignored_layers,

examples/offline_inference/text_to_video/text_to_video.md

@@ -231,7 +231,7 @@ python text_to_video.py \
 - `--frame-rate`: generation FPS for pipelines that require it (e.g., LTX2).
 - `--audio-sample-rate`: audio sample rate for embedded audio (when the
   pipeline returns audio; LTX-2.3 outputs 48kHz audio).
-- `--quantization`: quantization method (such as `fp8` for FP8).
+- `--quantization`: quantization method (`fp8` for FP8, `gguf` for GGUF).
 - `--flow-shift`: scheduler flow_shift parameter.
 - `--extra-body`: JSON object of model-specific generation params, filtered against the model's declared `extra_body_params` (see [`vllm_omni/model_extras`](../../../vllm_omni/model_extras)). Used by Cosmos3 (see above).
 

examples/offline_inference/text_to_video/text_to_video.py

@@ -227,7 +227,7 @@ def parse_args() -> argparse.Namespace:
         "--quantization",
         type=str,
         default=None,
-        choices=["fp8", "mxfp8", "mxfp4", "mxfp4_dualscale", "int8"],
+        choices=["fp8", "mxfp8", "mxfp4", "mxfp4_dualscale", "int8", "gguf"],
         help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 (NPU). mxfp4: W4A4 MXFP4 (NPU). mxfp4_dualscale: W4A4 MXFP4 dual-scale + BF16 fallback mixed (NPU). fp8: online FP8 (GPU).",
     )
 

examples/offline_inference/text_to_video/text_to_video_lingbot.py

@@ -0,0 +1,124 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from __future__ import annotations
+
+import argparse
+import time
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.outputs import OmniRequestOutput
+from vllm_omni.platforms import current_omni_platform
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Generate text-to-video with the LingBot dense preset.")
+    parser.add_argument("--model", default="/home/models/lingbot-video-dense-1.3b")
+    parser.add_argument("--prompt", default="a robotic arm picks up a red block")
+    parser.add_argument("--negative-prompt", default="")
+    parser.add_argument("--output", default="lingbot_video_output.mp4")
+    parser.add_argument("--height", type=int, default=192)
+    parser.add_argument("--width", type=int, default=320)
+    parser.add_argument("--num-frames", type=int, default=9)
+    parser.add_argument("--num-inference-steps", type=int, default=2)
+    parser.add_argument("--guidance-scale", type=float, default=3.0)
+    parser.add_argument("--flow-shift", type=float, default=3.0)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--fps", type=int, default=24)
+    parser.add_argument("--output-type", choices=["pt", "np"], default="pt")
+    parser.add_argument("--batch-cfg", action="store_true")
+    return parser.parse_args()
+
+
+def _peak_memory_mb(output: Any) -> float:
+    result = output[0] if isinstance(output, list) and output else output
+    return float(getattr(result, "peak_memory_mb", 0.0) or 0.0)
+
+
+def _extract_video(output: Any) -> Any:
+    if isinstance(output, list):
+        output = output[0] if output else None
+    if isinstance(output, OmniRequestOutput):
+        if output.final_output_type != "image":
+            raise ValueError(f"Unexpected output type {output.final_output_type!r}.")
+        if not output.images:
+            raise ValueError("No video frames found in OmniRequestOutput.")
+        return output.images[0] if len(output.images) == 1 else output.images
+    return output
+
+
+def _to_frame_array(video: Any) -> np.ndarray:
+    if isinstance(video, torch.Tensor):
+        tensor = video.detach().cpu()
+        if tensor.dim() == 5:
+            tensor = tensor[0] if tensor.shape[1] not in (3, 4) else tensor[0].permute(1, 2, 3, 0)
+        elif tensor.dim() == 4 and tensor.shape[0] in (3, 4):
+            tensor = tensor.permute(1, 2, 3, 0)
+        if tensor.is_floating_point():
+            if float(tensor.min()) < 0.0:
+                tensor = tensor.clamp(-1, 1) * 0.5 + 0.5
+            else:
+                tensor = tensor.clamp(0, 1)
+        return tensor.float().numpy()
+    if isinstance(video, np.ndarray):
+        array = video[0] if video.ndim == 5 else video
+        if np.issubdtype(array.dtype, np.integer):
+            return array.astype(np.float32) / 255.0
+        return array.astype(np.float32, copy=False)
+    if isinstance(video, list):
+        return np.stack([_to_frame_array(frame) for frame in video], axis=0)
+    return np.asarray(video, dtype=np.float32)
+
+
+def main() -> None:
+    args = parse_args()
+    generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(args.seed)
+    omni = Omni(
+        model=args.model,
+        model_class_name="LingBotVideoPipeline",
+        flow_shift=args.flow_shift,
+        parallel_config=DiffusionParallelConfig(),
+    )
+
+    prompt: dict[str, str] = {"prompt": args.prompt}
+    if args.negative_prompt:
+        prompt["negative_prompt"] = args.negative_prompt
+
+    sampling_params = OmniDiffusionSamplingParams(
+        height=args.height,
+        width=args.width,
+        num_frames=args.num_frames,
+        num_inference_steps=args.num_inference_steps,
+        guidance_scale=args.guidance_scale,
+        generator=generator,
+        output_type=args.output_type,
+        extra_args={"batch_cfg": args.batch_cfg, "shift": args.flow_shift},
+    )
+
+    start = time.perf_counter()
+    output = omni.generate(prompt, sampling_params)
+    elapsed = time.perf_counter() - start
+    print(f"Total generation time: {elapsed:.4f} seconds ({elapsed * 1000:.2f} ms)")
+    peak_mb = _peak_memory_mb(output)
+    if peak_mb:
+        print(f"Worker peak GPU memory (reserved): {peak_mb:.2f} MiB ({peak_mb / 1024:.2f} GiB)")
+
+    video = _to_frame_array(_extract_video(output))
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    from diffusers.utils import export_to_video
+
+    export_to_video(list(video), str(output_path), fps=args.fps)
+    print(f"Saved generated video to {output_path}")
+    omni.close()
+
+
+if __name__ == "__main__":
+    main()

recipes/README.md

@@ -44,11 +44,11 @@ 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 |
 | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
+| [`Robbyant/LingBot-Video.md`](./Robbyant/LingBot-Video.md) | Native dense text-to-video serving | 1x L20X / A100 / H100 |
 | [`cosmos3/Cosmos3-Nano.md`](./cosmos3/Cosmos3-Nano.md) | Text-to-image, text-to-video, image-to-video, video-to-video generation, text to video with sound, action policy | 1x H200 141GB / B300 |
 | [`cosmos3/Cosmos3-Super.md`](./cosmos3/Cosmos3-Super.md) | 64B T2I / T2V / I2V / V2V generation (+ optional audio) / Action policy | 8x H200/H100/A100 / 2x H200 / B300 |
 | [`OpenBMB/MiniCPM-o-4_5.md`](./OpenBMB/MiniCPM-o-4_5.md) | Online serving for omni multimodal chat (text / image / audio / video → text + 24 kHz speech) | 2x A100/H100 80GB / 3x mid-tier GPU / 8x RTX 4090 24GB |

recipes/Robbyant/LingBot-Video.md

@@ -0,0 +1,143 @@
+# LingBot-Video
+
+> Native dense text-to-video serving for `robbyant/lingbot-video-dense-1.3b`
+
+## Summary
+
+- Vendor: Robbyant
+- Model: `robbyant/lingbot-video-dense-1.3b`
+- Task: Text-to-video generation
+- Mode: Offline generation and online serving with the OpenAI-compatible `/v1/videos` API
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe when you want to run the dense LingBot-Video checkpoint with
+vLLM-Omni's native pipeline. The runtime path does not import the upstream
+`lingbot_video` Python package; it loads the checkpoint components directly
+with the in-tree `LingBotVideoPipeline`, dense DiT transformer, shared FlowUniPC
+scheduler, Qwen3-VL text encoder, and Wan VAE.
+
+This recipe is intentionally text-to-video only. Text-to-image, image-conditioned
+video, TI2V, MoE checkpoints, and fused-expert kernels are not covered here.
+
+## References
+
+- Dense checkpoint: <https://huggingface.co/robbyant/lingbot-video-dense-1.3b>
+- Upstream project: <https://github.com/Robbyant/lingbot-video>
+- Related offline example: [`examples/offline_inference/text_to_video/text_to_video_lingbot.py`](../../examples/offline_inference/text_to_video/text_to_video_lingbot.py)
+- Related online video API docs: [`docs/serving/videos_api.md`](../../docs/serving/videos_api.md)
+
+## Hardware Support
+
+This recipe documents the CUDA single-GPU dense checkpoint path. Multi-GPU
+parallelism, Cache-DiT, CPU offload, and the LingBot MoE path are not validated
+for this model in this PR.
+
+## GPU
+
+### 1 x NVIDIA L20X / A100 / H100
+
+The dense 1.3B checkpoint has been smoke-tested on a single NVIDIA L20X at a
+small validation shape (`192x320`, 9 frames, 2 steps). Use a larger GPU or lower
+resolution/frame count if your target workload runs out of memory.
+
+#### Offline T2V
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python examples/offline_inference/text_to_video/text_to_video_lingbot.py \
+  --model robbyant/lingbot-video-dense-1.3b \
+  --prompt "a robotic arm picks up a red block" \
+  --output lingbot_t2v.mp4 \
+  --height 192 \
+  --width 320 \
+  --num-frames 9 \
+  --num-inference-steps 2 \
+  --guidance-scale 3.0 \
+  --flow-shift 3.0 \
+  --seed 42 \
+  --fps 24
+```
+
+#### Online Serving
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+vllm serve robbyant/lingbot-video-dense-1.3b \
+  --omni \
+  --model-class-name LingBotVideoPipeline \
+  --port 8091
+```
+
+After the server is ready, submit a text-to-video job:
+
+```bash
+create_response=$(curl -s http://localhost:8091/v1/videos \
+  -F "model=robbyant/lingbot-video-dense-1.3b" \
+  -F "prompt=a robotic arm picks up a red block" \
+  -F "width=320" \
+  -F "height=192" \
+  -F "num_frames=9" \
+  -F "fps=24" \
+  -F "num_inference_steps=2" \
+  -F "guidance_scale=3.0" \
+  -F "flow_shift=3.0" \
+  -F "seed=42")
+
+video_id=$(echo "${create_response}" | jq -r '.id')
+while true; do
+  status=$(curl -s "http://localhost:8091/v1/videos/${video_id}" | jq -r '.status')
+  if [ "${status}" = "completed" ]; then
+    break
+  fi
+  if [ "${status}" = "failed" ]; then
+    curl -s "http://localhost:8091/v1/videos/${video_id}" | jq .
+    exit 1
+  fi
+  sleep 2
+done
+
+curl -L "http://localhost:8091/v1/videos/${video_id}/content" -o lingbot_t2v.mp4
+```
+
+## Key Parameters
+
+| Parameter | Suggested smoke value | Notes |
+|-----------|-----------------------|-------|
+| `height` | `192` | Must be a multiple of 16 |
+| `width` | `320` | Must be a multiple of 16 |
+| `num_frames` | `9` | Must be `1` or `4n + 1`; this PR validates T2V with video outputs |
+| `num_inference_steps` | `2` | Use more steps for quality sweeps |
+| `guidance_scale` | `3.0` | CFG is active when this is greater than `1.0` |
+| `flow_shift` | `3.0` | Scheduler flow-shift; aliases the pipeline's internal `shift` |
+| `negative_prompt` | model default | Optional text describing artifacts to avoid |
+| `fps` | `24` | Output MP4 frame rate |
+
+## Validation
+
+Local dense smoke run:
+
+- Shape: 9 frames at `192x320`
+- Steps: 2
+- Request generation time: `0.2923s`
+- Peak reserved GPU memory: `14548 MiB`
+
+Local parity harness against the upstream repository:
+
+- Shape: `[9, 192, 320, 3]`
+- MAE: `0.0065238`
+- MSE: `0.00006650`
+- PSNR: `41.77 dB`
+- Native request time: `0.2875s`
+
+Do not treat these as production benchmarks; they are draft-PR validation
+numbers for the small dense smoke configuration.
+
+## Known Limitations
+
+- T2V only. T2I, I2V, and TI2V are not claimed by this PR.
+- Dense checkpoint only. MoE/fused-expert support belongs in a separate PR.
+- No Cache-DiT, TeaCache, tensor parallelism, sequence parallelism, CFG
+  parallelism, HSDP, CPU offload, or VAE patch parallelism validation yet.
+- Only one request per LingBot pipeline batch is currently supported.

recipes/krea/Krea-2.md

@@ -1,143 +0,0 @@
-# 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.

tests/ar_diffusion/test_kv_cache.py

@@ -9,9 +9,14 @@ KV stack on CPU (block bookkeeping only, no GPU tensors).
 import pytest
 import torch
 from vllm.v1.kv_cache_interface import KVCacheSpecKind, get_kv_cache_spec_kind
-from vllm.v1.kv_cache_spec_registry import KVCacheSpecRegistry
 from vllm.v1.request import RequestStatus
 
+try:
+    from vllm.v1.kv_cache_spec_registry import KVCacheSpecRegistry
+except ModuleNotFoundError:
+    KVCacheSpecRegistry = None
+    from vllm.v1.core.single_type_kv_cache_manager import spec_manager_map
+
 from vllm_omni.experimental.ar_diffusion.kv_cache import (
     ARDiffusionKVCache,
     ARDiffusionKVConfig,
@@ -48,7 +53,10 @@ def test_spec_registration_resolves_to_chunk_window_manager():
     # Without explicit registration the MRO walk would fall back to the parent
     # SlidingWindowManager; assert the subclass manager wins.
     spec = make_spec()
-    assert KVCacheSpecRegistry.get_manager_class(spec) is ChunkWindowManager
+    if KVCacheSpecRegistry is not None:
+        assert KVCacheSpecRegistry.get_manager_class(spec) is ChunkWindowManager
+    else:
+        assert spec_manager_map[type(spec)] is ChunkWindowManager
 
 
 def test_spec_kind_is_sliding_window():

tests/config/composable_parallel/test_apply.py

@@ -1,154 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Tests for applying strategy specs onto merged stage configs."""
-
-from __future__ import annotations
-
-import pytest
-
-from vllm_omni.config.composable_parallel import (
-    Broadcast,
-    FanInByStage,
-    MeshAxisSpec,
-    RouteByStage,
-    StrategyApplyError,
-    StrategySpec,
-    TakeRank,
-    apply_strategy_specs,
-)
-from vllm_omni.config.stage_config import StageConfig
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def _tp(size: int) -> StrategySpec:
-    return StrategySpec("tp", MeshAxisSpec("tp", size), Broadcast(), TakeRank())
-
-
-def _stage_replica(size: int, policy: str = "round_robin") -> StrategySpec:
-    return StrategySpec("stage_replica", MeshAxisSpec("stage_replica", size), RouteByStage(policy), FanInByStage())
-
-
-def _stage(stage_id: int, model_stage: str, engine_args=None, runtime=None) -> StageConfig:
-    return StageConfig(
-        stage_id=stage_id,
-        model_stage=model_stage,
-        yaml_engine_args=dict(engine_args or {}),
-        yaml_runtime=dict(runtime or {"num_replicas": 1}),
-    )
-
-
-def _qwen_stages() -> list[StageConfig]:
-    return [
-        _stage(0, "thinker"),
-        _stage(1, "talker"),
-        _stage(2, "code2wav"),
-    ]
-
-
-def test_apply_tp_by_role():
-    stages = _qwen_stages()
-    apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-    assert stages[0].yaml_engine_args["tensor_parallel_size"] == 2
-    # untouched roles keep their config
-    assert "tensor_parallel_size" not in stages[1].yaml_engine_args
-
-
-def test_apply_by_model_stage():
-    stages = _qwen_stages()
-    apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-    assert stages[0].yaml_engine_args["tensor_parallel_size"] == 2
-
-
-def test_apply_stage_replica_sets_num_replicas_and_surfaces_lb():
-    stages = _qwen_stages()
-    result = apply_strategy_specs(stages, {"talker": [_stage_replica(2, "round_robin")]})
-    assert stages[1].yaml_runtime["num_replicas"] == 2
-    assert result.omni_lb_policy == "round-robin"
-
-
-def test_only_declared_axes_are_written():
-    stages = _qwen_stages()
-    # strategy declares only stage_replica -> tp must not be forced.
-    apply_strategy_specs(stages, {"talker": [_stage_replica(2)]})
-    assert "tensor_parallel_size" not in stages[1].yaml_engine_args
-
-
-def test_conflict_on_explicit_tp():
-    stages = _qwen_stages()
-    stages[0].yaml_engine_args["tensor_parallel_size"] = 4
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-
-
-def test_equal_explicit_value_is_noop():
-    stages = _qwen_stages()
-    stages[0].yaml_engine_args["tensor_parallel_size"] = 2
-    apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-    assert stages[0].yaml_engine_args["tensor_parallel_size"] == 2
-
-
-def test_explicit_none_conflicts_with_derived_value():
-    # An explicit YAML ``null`` (``tensor_parallel_size: null``) is a *present*
-    # value, not a missing key, so a strategy deriving a non-None size must raise
-    # rather than silently clobber the explicit None.
-    stages = _qwen_stages()
-    stages[0].yaml_engine_args["tensor_parallel_size"] = None
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-
-
-def test_missing_key_is_filled_without_conflict():
-    # A genuinely absent key (never set in the YAML) is filled by the strategy
-    # and must NOT raise — the contrast case to an explicit None.
-    stages = _qwen_stages()
-    assert "tensor_parallel_size" not in stages[0].yaml_engine_args
-    apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-    assert stages[0].yaml_engine_args["tensor_parallel_size"] == 2
-
-
-def test_num_replicas_conflict():
-    stages = _qwen_stages()
-    stages[1].yaml_runtime["num_replicas"] = 3
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(stages, {"talker": [_stage_replica(2)]})
-
-
-def test_device_count_ok_template():
-    stages = _qwen_stages()
-    stages[0].yaml_runtime["devices"] = "0,1"
-    apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-    assert stages[0].yaml_engine_args["tensor_parallel_size"] == 2
-
-
-def test_device_count_ok_pool():
-    # tp=2 -> world=2; 2 replicas -> pool of 4 device ids is valid.
-    stages = _qwen_stages()
-    stages[1].yaml_runtime["devices"] = "0,1,2,3"
-    apply_strategy_specs(stages, {"talker": [_tp(2), _stage_replica(2)]})
-    assert stages[1].yaml_runtime["num_replicas"] == 2
-
-
-def test_device_count_mismatch():
-    stages = _qwen_stages()
-    stages[0].yaml_runtime["devices"] = "0,1,2"
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-
-
-def test_unknown_role_raises():
-    stages = _qwen_stages()
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(stages, {"nonexistent": [_tp(2)]})
-
-
-def test_conflicting_lb_policy_across_roles():
-    stages = _qwen_stages()
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(
-            stages,
-            {
-                "talker": [_stage_replica(2, "round_robin")],
-                "code2wav": [_stage_replica(2, "least_queue")],
-            },
-        )

tests/config/composable_parallel/test_factory_integration.py

@@ -1,141 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Integration tests: strategy overlay on the real merge_pipeline_deploy seam."""
-
-from __future__ import annotations
-
-import logging
-from pathlib import Path
-
-import pytest
-
-from vllm_omni.config.composable_parallel import (
-    Broadcast,
-    FanInByStage,
-    MeshAxisSpec,
-    RouteByStage,
-    StrategyApplyError,
-    StrategySpec,
-    TakeRank,
-    apply_strategy_specs,
-)
-from vllm_omni.config.config_factory import StageConfigFactory
-from vllm_omni.config.pipeline_registry import OMNI_PIPELINES
-from vllm_omni.config.stage_config import load_deploy_config, merge_pipeline_deploy
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-_DEPLOY = Path(__file__).parents[3] / "vllm_omni" / "deploy" / "qwen2_5_omni.yaml"
-
-
-def _tp(size: int) -> StrategySpec:
-    return StrategySpec("tp", MeshAxisSpec("tp", size), Broadcast(), TakeRank())
-
-
-def _stage_replica(size: int, policy: str = "round_robin") -> StrategySpec:
-    return StrategySpec("stage_replica", MeshAxisSpec("stage_replica", size), RouteByStage(policy), FanInByStage())
-
-
-def _qwen_stages():
-    pipeline = OMNI_PIPELINES["qwen2_5_omni"]
-    deploy = load_deploy_config(_DEPLOY)
-    return merge_pipeline_deploy(pipeline, deploy)
-
-
-def _stage(stages, role):
-    return next(s for s in stages if s.model_stage == role)
-
-
-def test_overlay_tp_on_thinker():
-    stages = _qwen_stages()
-    # The bundled deploy pins the thinker to one GPU; a TP=2 strategy needs a
-    # matching 2-GPU layout (mirrors what a TP2 deploy would declare).
-    _stage(stages, "thinker").yaml_runtime["devices"] = "0,1"
-    result = apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-    assert _stage(result.stages, "thinker").yaml_engine_args["tensor_parallel_size"] == 2
-
-
-def test_device_guard_rejects_tp_on_single_gpu_deploy():
-    # The bundled deploy pins the thinker to a single GPU, so the pre-spawn
-    # device check must refuse a TP=2 strategy on it.
-    stages = _qwen_stages()
-    with pytest.raises(StrategyApplyError):
-        apply_strategy_specs(stages, {"thinker": [_tp(2)]})
-
-
-def test_overlay_stage_replica_on_talker():
-    stages = _qwen_stages()
-    result = apply_strategy_specs(stages, {"talker": [_stage_replica(2, "round_robin")]})
-    assert _stage(result.stages, "talker").yaml_runtime["num_replicas"] == 2
-    assert result.omni_lb_policy == "round-robin"
-
-
-def test_overlay_mixed_roles():
-    stages = _qwen_stages()
-    _stage(stages, "thinker").yaml_runtime["devices"] = "0,1"
-    result = apply_strategy_specs(
-        stages,
-        {
-            "thinker": [_tp(2)],
-            "talker": [_stage_replica(2, "round_robin")],
-            "code2wav": [_stage_replica(2, "round_robin")],
-        },
-    )
-    by_role = {s.model_stage: s for s in result.stages}
-    assert by_role["thinker"].yaml_engine_args["tensor_parallel_size"] == 2
-    assert by_role["talker"].yaml_runtime["num_replicas"] == 2
-    assert by_role["code2wav"].yaml_runtime["num_replicas"] == 2
-    assert result.omni_lb_policy == "round-robin"
-
-
-def _resolved(stages, role):
-    """Return a role's fully-resolved (post-CLI) OmegaConf stage config."""
-    stage = next(s for s in stages if s.model_stage == role)
-    return stage.to_omegaconf()
-
-
-def test_device_check_survives_cli_override():
-    # Strategy replicates the talker (1-GPU template -> valid at apply time),
-    # but a CLI --stage_1_devices with 3 ids must NOT slip past the device
-    # guard: effective world=1, replicas=2 admits only 1 or 2 device ids.
-    pipeline_cfg = OMNI_PIPELINES["qwen2_5_omni"]
-    with pytest.raises(StrategyApplyError):
-        StageConfigFactory._create_from_registry(
-            "qwen2_5_omni",
-            pipeline_cfg,
-            cli_overrides={"stage_1_devices": "0,1,2"},
-            strategy_specs={"talker": [_stage_replica(2, "round_robin")]},
-        )
-
-
-def test_cli_overrides_strategy_with_warning():
-    # Strategy derives num_replicas=2 for the talker; a CLI override to 3 wins
-    # (it is the most explicit user action) but must be surfaced loudly rather
-    # than silently. The resulting layout (1 template device) stays valid.
-    #
-    # vLLM's logger sets propagate=False, so attach a handler directly to it
-    # rather than relying on pytest's caplog (which listens on the root logger).
-    messages: list[str] = []
-
-    class _Capture(logging.Handler):
-        def emit(self, record: logging.LogRecord) -> None:
-            messages.append(record.getMessage())
-
-    log = logging.getLogger("vllm_omni.config.config_factory")
-    handler = _Capture(level=logging.WARNING)
-    log.addHandler(handler)
-    try:
-        pipeline_cfg = OMNI_PIPELINES["qwen2_5_omni"]
-        stages, _ = StageConfigFactory._create_from_registry(
-            "qwen2_5_omni",
-            pipeline_cfg,
-            cli_overrides={"stage_1_num_replicas": 3},
-            strategy_specs={"talker": [_stage_replica(2, "round_robin")]},
-        )
-    finally:
-        log.removeHandler(handler)
-
-    # CLI value wins in the resolved config.
-    assert _resolved(stages, "talker").runtime.num_replicas == 3
-    # ...and the override was warned about, naming the conflicting field.
-    assert any("num_replicas" in m and "overrides the strategy-derived" in m for m in messages)

tests/config/composable_parallel/test_spec_types.py

@@ -1,42 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Tests for the vendored spec/pattern data types."""
-
-from __future__ import annotations
-
-import pytest
-
-from vllm_omni.config.composable_parallel.aggregation import AggregationPattern, TakeRank
-from vllm_omni.config.composable_parallel.routing import RoutingPattern
-from vllm_omni.config.composable_parallel.spec import MeshAxisSpec
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def test_mesh_axis_rejects_bad_kind():
-    with pytest.raises(ValueError):
-        MeshAxisSpec(kind="not_a_kind", size=2)  # type: ignore[arg-type]
-
-
-def test_mesh_axis_rejects_nonpositive_size():
-    with pytest.raises(ValueError):
-        MeshAxisSpec(kind="tp", size=0)
-
-
-def test_routing_pattern_is_closed():
-    with pytest.raises(TypeError):
-
-        class RogueRouting(RoutingPattern):  # noqa: F811
-            pass
-
-
-def test_aggregation_pattern_is_closed():
-    with pytest.raises(TypeError):
-
-        class RogueAggregation(AggregationPattern):  # noqa: F811
-            pass
-
-
-def test_take_rank_type_guard():
-    with pytest.raises(TypeError):
-        TakeRank(rank=True)  # bool is not a valid rank

tests/config/composable_parallel/test_strategy_loader.py

@@ -1,97 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Tests for parsing/loading strategy files."""
-
-from __future__ import annotations
-
-import pytest
-
-from vllm_omni.config.composable_parallel.routing import Broadcast, RouteByStage
-from vllm_omni.config.composable_parallel.strategy_loader import (
-    StrategyLoadError,
-    load_strategy_specs,
-    parse_strategy_specs,
-)
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def test_parse_basic():
-    data = {
-        "strategies": {
-            "thinker": [{"axis": "tp", "size": 2}],
-            "talker": [{"axis": "stage_replica", "size": 2, "routing": "round_robin"}],
-        }
-    }
-    specs = parse_strategy_specs(data)
-    assert set(specs) == {"thinker", "talker"}
-
-    (tp_spec,) = specs["thinker"]
-    assert tp_spec.mesh_axis.kind == "tp"
-    assert tp_spec.mesh_axis.size == 2
-    assert isinstance(tp_spec.routing, Broadcast)
-
-    (sr_spec,) = specs["talker"]
-    assert sr_spec.mesh_axis.kind == "stage_replica"
-    assert isinstance(sr_spec.routing, RouteByStage)
-    assert sr_spec.routing.routing_policy == "round_robin"
-
-
-def test_parse_without_strategies_key():
-    # A bare role mapping (no top-level "strategies") is also accepted.
-    specs = parse_strategy_specs({"thinker": [{"axis": "tp", "size": 4}]})
-    assert specs["thinker"][0].mesh_axis.size == 4
-
-
-def test_l1_owner_goes_to_shard_extension():
-    specs = parse_strategy_specs({"talker": [{"axis": "stage_replica", "size": 2, "l1_owner": "delegated"}]})
-    assert specs["talker"][0].shard_extension["l1_owner"] == "delegated"
-
-
-def test_missing_axis_raises():
-    with pytest.raises(StrategyLoadError):
-        parse_strategy_specs({"thinker": [{"size": 2}]})
-
-
-def test_missing_size_raises():
-    with pytest.raises(StrategyLoadError):
-        parse_strategy_specs({"thinker": [{"axis": "tp"}]})
-
-
-def test_routing_on_non_policy_axis_raises():
-    with pytest.raises(StrategyLoadError):
-        parse_strategy_specs({"thinker": [{"axis": "tp", "size": 2, "routing": "random"}]})
-
-
-def test_bad_size_raises():
-    with pytest.raises(StrategyLoadError):
-        parse_strategy_specs({"thinker": [{"axis": "tp", "size": "two"}]})
-
-
-def test_entries_must_be_list():
-    with pytest.raises(StrategyLoadError):
-        parse_strategy_specs({"thinker": {"axis": "tp", "size": 2}})
-
-
-def test_non_mapping_entry_raises():
-    # A list whose elements are not mappings (e.g. a bare string) must raise a
-    # StrategyLoadError, not an opaque TypeError from dict(entry).
-    with pytest.raises(StrategyLoadError):
-        parse_strategy_specs({"thinker": ["tp"]})
-
-
-def test_load_from_file(tmp_path):
-    path = tmp_path / "strategy.yaml"
-    path.write_text(
-        "strategies:\n"
-        "  thinker:\n"
-        "    - axis: tp\n"
-        "      size: 2\n"
-        "  talker:\n"
-        "    - axis: stage_replica\n"
-        "      size: 2\n"
-        "      routing: least_queue\n"
-    )
-    specs = load_strategy_specs(str(path))
-    assert specs["thinker"][0].mesh_axis.size == 2
-    assert specs["talker"][0].routing.routing_policy == "least_queue"

tests/config/composable_parallel/test_translator.py

@@ -1,195 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Tests for StrategySpec -> OmniParallelConfig translation."""
-
-from __future__ import annotations
-
-import dataclasses
-
-import pytest
-
-from vllm_omni.config.composable_parallel import (
-    AxisTranslationError,
-    Broadcast,
-    FanInByStage,
-    MeshAxisSpec,
-    PartitionByHash,
-    PipelineMicrobatch,
-    RouteByStage,
-    StrategySpec,
-    TakeRank,
-    Union,
-    translate_strategy_stack,
-)
-from vllm_omni.config.composable_parallel.aggregation import StitchPipeline
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def _tp(size: int) -> StrategySpec:
-    return StrategySpec("tp", MeshAxisSpec("tp", size), Broadcast(), TakeRank())
-
-
-def _dp(size: int, policy: str = "round_robin") -> StrategySpec:
-    return StrategySpec("dp", MeshAxisSpec("dp", size), RouteByStage(policy), Union())
-
-
-def _pp(size: int) -> StrategySpec:
-    return StrategySpec("pp", MeshAxisSpec("pp", size), PipelineMicrobatch(), StitchPipeline())
-
-
-def _ep(size: int) -> StrategySpec:
-    return StrategySpec("ep", MeshAxisSpec("ep", size), Broadcast(), Union())
-
-
-def _stage_replica(size: int, policy: str = "round_robin") -> StrategySpec:
-    return StrategySpec("stage_replica", MeshAxisSpec("stage_replica", size), RouteByStage(policy), FanInByStage())
-
-
-def test_tp_only():
-    cfg = translate_strategy_stack([_tp(2)])
-    assert cfg.tensor_parallel_size == 2
-    assert cfg.data_parallel_size == 1
-    assert cfg.world_size == 2
-    assert cfg.stage_replica_size == 1
-    assert cfg.l1_owners == {"tp": "engine"}
-
-
-def test_dp_is_engine_owned():
-    cfg = translate_strategy_stack([_dp(2)])
-    assert cfg.data_parallel_size == 2
-    assert cfg.world_size == 2
-    # engine DP does not emit an omni LB policy (vLLM's internal DP LB is used).
-    assert cfg.omni_lb_policy is None
-    assert cfg.as_engine_kwargs().get("omni_lb_policy") is None
-    assert cfg.l1_owners["dp"] == "engine"
-
-
-def test_tp_times_dp_world_size():
-    cfg = translate_strategy_stack([_tp(2), _dp(2)])
-    assert cfg.world_size == 4
-    assert cfg.tensor_parallel_size == 2
-    assert cfg.data_parallel_size == 2
-
-
-def test_stage_replica_is_delegated():
-    cfg = translate_strategy_stack([_stage_replica(3, "least_queue")])
-    # stage_replica is NOT a world dimension.
-    assert cfg.world_size == 1
-    assert cfg.stage_replica_size == 3
-    assert cfg.omni_lb_policy == "least-queue-length"
-    assert cfg.l1_owners["stage_replica"] == "delegated"
-    # stage_replica_size is not an engine kwarg; it's a per-stage deploy knob.
-    assert "stage_replica_size" not in cfg.as_engine_kwargs()
-
-
-def test_stage_replica_lb_policy_only_when_replicated():
-    cfg = translate_strategy_stack([_stage_replica(1)])
-    assert cfg.stage_replica_size == 1
-    assert cfg.as_engine_kwargs().get("omni_lb_policy") is None
-
-
-def test_tp_with_stage_replica():
-    cfg = translate_strategy_stack([_tp(2), _stage_replica(2)])
-    assert cfg.world_size == 2  # tp only
-    assert cfg.tensor_parallel_size == 2
-    assert cfg.stage_replica_size == 2
-    assert cfg.omni_lb_policy == "round-robin"
-
-
-def test_ep_must_match_tp_dp_product():
-    cfg = translate_strategy_stack([_tp(2), _ep(2)])
-    assert cfg.enable_expert_parallel is True
-    assert cfg.as_engine_kwargs()["enable_expert_parallel"] is True
-
-
-def test_ep_mismatch_raises():
-    with pytest.raises(Exception) as exc:
-        translate_strategy_stack([_tp(2), _ep(4)])
-    assert "must equal" in str(exc.value)
-
-
-def test_duplicate_kind_rejected():
-    with pytest.raises(AxisTranslationError):
-        translate_strategy_stack([_tp(2), _tp(2)])
-
-
-def test_unsupported_kind_rejected():
-    spec = StrategySpec("sp", MeshAxisSpec("sp_ulysses", 2), Broadcast(), TakeRank())
-    with pytest.raises(AxisTranslationError):
-        translate_strategy_stack([spec])
-
-
-def test_dp_hash_routing_rejected():
-    spec = StrategySpec("dp", MeshAxisSpec("dp", 2), PartitionByHash(), Union())
-    with pytest.raises(NotImplementedError):
-        translate_strategy_stack([spec])
-
-
-def test_dp_invalid_routing_policy_rejected():
-    spec = StrategySpec("dp", MeshAxisSpec("dp", 2), RouteByStage("bogus"), Union())
-    with pytest.raises(AxisTranslationError):
-        translate_strategy_stack([spec])
-
-
-def test_lb_policy_never_in_engine_kwargs_even_when_replicated():
-    # omni_lb_policy is pipeline-wide (applied at engine construction), not a
-    # per-stage engine arg, so it must never leak into as_engine_kwargs().
-    cfg = translate_strategy_stack([_tp(2), _stage_replica(2, "least_queue")])
-    assert cfg.omni_lb_policy == "least-queue-length"
-    assert "omni_lb_policy" not in cfg.as_engine_kwargs()
-
-
-def test_stage_replica_hash_routing_rejected():
-    spec = StrategySpec("sr", MeshAxisSpec("stage_replica", 2), RouteByStage("hash"), FanInByStage())
-    with pytest.raises(NotImplementedError):
-        translate_strategy_stack([spec])
-
-
-def test_dp_wrong_owner_rejected():
-    spec = StrategySpec(
-        "dp",
-        MeshAxisSpec("dp", 2),
-        RouteByStage("round_robin"),
-        Union(),
-        shard_extension={"l1_owner": "delegated"},
-    )
-    with pytest.raises(AxisTranslationError):
-        translate_strategy_stack([spec])
-
-
-def test_stage_replica_wrong_owner_rejected():
-    spec = StrategySpec(
-        "sr",
-        MeshAxisSpec("stage_replica", 2),
-        RouteByStage("round_robin"),
-        FanInByStage(),
-        shard_extension={"l1_owner": "engine"},
-    )
-    with pytest.raises(AxisTranslationError):
-        translate_strategy_stack([spec])
-
-
-@pytest.mark.parametrize(
-    "specs",
-    [
-        [_tp(2)],
-        [_dp(2)],
-        [_tp(2), _dp(2)],
-        [_tp(2), _ep(2)],
-        [_tp(2), _stage_replica(2)],
-        [_pp(2)],
-    ],
-)
-def test_as_engine_kwargs_are_valid_omni_engine_args(specs):
-    # Guard against desync: every key emitted by as_engine_kwargs() must be a
-    # real OmniEngineArgs field, otherwise constructing the per-stage engine args
-    # would explode at runtime. Keep this CPU-only by checking field names rather
-    # than instantiating the engine.
-    from vllm_omni.engine.arg_utils import OmniEngineArgs
-
-    valid_fields = {f.name for f in dataclasses.fields(OmniEngineArgs)}
-    cfg = translate_strategy_stack(specs)
-    kwargs = cfg.as_engine_kwargs()
-    unknown = set(kwargs) - valid_fields
-    assert not unknown, f"as_engine_kwargs emitted non-OmniEngineArgs fields: {unknown}"

tests/config/test_omni_config.py

@@ -751,14 +751,3 @@ def test_diffusion_config_field_classification_covers_current_fields():
         "distributed_executor_backend",
     } <= omni_config_module._DIFFUSION_SHARED_CONFIG_FIELDS
     assert "prompt_file_path" in omni_config_module._DIFFUSION_RUNTIME_CONFIG_FIELDS
-
-
-def test_diffusion_config_projection_keeps_mapping_quantization_config_serializable():
-    quantization_config = {
-        "method": "example_quant",
-        "weights": "weights.bin",
-    }
-
-    cfg = omni_config_module._DiffusionConfigProjection.from_kwargs(quantization_config=quantization_config)
-
-    assert cfg.quantization_config == quantization_config

tests/dfx/perf/tests/test_lingbot_video_vllm_omni.json

@@ -0,0 +1,34 @@
+[
+    {
+        "test_name": "test_lingbot_video_single_device",
+        "description": "Single-device LingBot dense T2V baseline at 320x192, 9 frames, 2 steps.",
+        "server_type": "vllm-omni",
+        "server_params": {
+            "model": "robbyant/lingbot-video-dense-1.3b",
+            "serve_args": {
+                "model-class-name": "LingBotVideoPipeline",
+                "enable-diffusion-pipeline-profiler": true
+            }
+        },
+        "benchmark_params": [
+            {
+                "name": "320x192_frames9_steps2",
+                "dataset": "random",
+                "task": "t2v",
+                "backend": "v1/videos",
+                "width": 320,
+                "height": 192,
+                "num-frames": 9,
+                "fps": 24,
+                "num-inference-steps": 2,
+                "num-prompts": 3,
+                "max-concurrency": 1,
+                "warmup-requests": 1,
+                "warmup-concurrency": 1,
+                "warmup-num-inference-steps": 2,
+                "enable-negative-prompt": true,
+                "skip-performance-assertion": true
+            }
+        ]
+    }
+]

tests/dfx/perf/tests/test_voxcpm2.json

@@ -3,10 +3,7 @@
     "test_name": "test_voxcpm2",
     "server_params": {
       "model": "openbmb/VoxCPM2",
-      "trust_remote_code": true,
+      "trust_remote_code": true
-      "extra_cli_args": [
-        "--trust-remote-code"
-      ]
     },
     "benchmark_params": [
       {

tests/dfx/reliability/invalid_param_test/test_invalid_image_editing.py

@@ -62,8 +62,8 @@ def _finalize_edits_form_data(template: dict[str, Any], server_model: str) -> di
         ),
         pytest.param(
             True,
-            {"prompt": "edit", "model": _IMAGE_EDITS_SERVER_MODEL, "layers": "1"},
+            {"prompt": "edit", "model": _IMAGE_EDITS_SERVER_MODEL, "layers": "2"},
-            ("Invalid layers", "2", "10"),
+            ("Invalid layers", "3", "1"),
             400,
             id="invalid_layers",
         ),

tests/dfx/reliability/invalid_param_test/test_invalid_image_generation.py

@@ -87,8 +87,8 @@ def _minimal_images_gen_json(omni_server: OmniServer) -> dict[str, object]:
         pytest.param(
             {"true_cfg_scale": 21.0}, ("true_cfg_scale", "less_than_equal", "20"), id="true_cfg_scale_above_max"
         ),
-        pytest.param({"layers": 1}, ("layers", "value_error", "2", "10"), id="layers_below_min"),
+        pytest.param({"layers": 2}, ("layers", "value_error", "3", "10"), id="layers_below_min"),
-        pytest.param({"layers": 11}, ("layers", "value_error", "2", "10"), id="layers_above_max"),
+        pytest.param({"layers": 11}, ("layers", "value_error", "3", "10"), id="layers_above_max"),
         pytest.param({"seed": -1}, ("seed", "greater_than_equal", "0"), id="seed_negative", marks=_SKIP_ISSUE_3649),
         pytest.param(
             {"seed": 2**32},

tests/diffusion/model_loader/gguf_adapters/test_flux2_klein_gguf_adapter.py

@@ -0,0 +1,112 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for the Flux2-Klein GGUF adapter."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.model_loader.gguf_adapters import get_gguf_adapter
+from vllm_omni.diffusion.model_loader.gguf_adapters.flux2_klein import (
+    Flux2KleinGGUFAdapter,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+def _make_od_config(
+    model_class_name: str = "Flux2Pipeline",
+    model_type: str = "flux",
+):
+    return SimpleNamespace(
+        model_class_name=model_class_name,
+        tf_model_config={"model_type": model_type},
+    )
+
+
+def _make_source(prefix: str = "", subfolder: str = "transformer"):
+    return SimpleNamespace(prefix=prefix, subfolder=subfolder)
+
+
+def test_flux2_klein_adapter_selected_for_flux_family():
+    adapter = get_gguf_adapter(
+        "dummy.gguf",
+        object(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    assert isinstance(adapter, Flux2KleinGGUFAdapter)
+
+
+def test_flux2_klein_adapter_matches_flux_model_type():
+    assert Flux2KleinGGUFAdapter.is_compatible(
+        _make_od_config(model_class_name="OtherPipeline", model_type="flux-dev"),
+        object(),
+        _make_source(),
+    )
+
+
+def test_flux2_klein_adapter_renames_core_projection_paths(monkeypatch: pytest.MonkeyPatch):
+    import vllm_omni.diffusion.model_loader.gguf_adapters.flux2_klein as flux_module
+
+    monkeypatch.setattr(
+        flux_module,
+        "gguf_quant_weights_iterator",
+        lambda _path: iter(
+            [
+                ("img_in.qweight", torch.ones((1, 1))),
+                ("time_in.in_layer.qweight_type", torch.tensor(1)),
+                ("double_blocks.0.img_attn.norm.query_norm.weight", torch.full((1, 1), 2.0)),
+                ("final_layer.linear.qweight", torch.full((1, 1), 3.0)),
+            ]
+        ),
+    )
+
+    adapter = Flux2KleinGGUFAdapter(
+        "dummy.gguf",
+        object(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    weights = list(adapter.weights_iterator())
+    names = [name for name, _ in weights]
+
+    assert "x_embedder.qweight" in names
+    assert "time_guidance_embed.timestep_embedder.linear_1.qweight_type" in names
+    assert "transformer_blocks.0.attn.norm_q.weight" in names
+    assert "proj_out.qweight" in names
+
+
+def test_flux2_klein_adapter_swaps_final_adaln_shift_and_scale(monkeypatch: pytest.MonkeyPatch):
+    import vllm_omni.diffusion.model_loader.gguf_adapters.flux2_klein as flux_module
+
+    monkeypatch.setattr(
+        flux_module,
+        "gguf_quant_weights_iterator",
+        lambda _path: iter(
+            [
+                (
+                    "final_layer.adaLN_modulation.1.scale",
+                    torch.tensor([1.0, 2.0, 3.0, 4.0]),
+                )
+            ]
+        ),
+    )
+
+    adapter = Flux2KleinGGUFAdapter(
+        "dummy.gguf",
+        object(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    weights = list(adapter.weights_iterator())
+
+    assert len(weights) == 1
+    assert weights[0][0] == "norm_out.linear.weight"
+    assert torch.equal(weights[0][1], torch.tensor([3.0, 4.0, 1.0, 2.0]))

tests/diffusion/model_loader/gguf_adapters/test_qwen_image_gguf_adapter.py

@@ -0,0 +1,184 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+
+from vllm_omni.diffusion.model_loader.gguf_adapters import get_gguf_adapter
+from vllm_omni.diffusion.model_loader.gguf_adapters.qwen_image import (
+    QwenImageGGUFAdapter,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+def _make_od_config(
+    model_class_name: str = "QwenImagePipeline",
+    model_type: str = "qwen_image",
+):
+    return SimpleNamespace(
+        model_class_name=model_class_name,
+        tf_model_config={"model_type": model_type},
+    )
+
+
+def _make_source(prefix: str = "", subfolder: str = "transformer"):
+    return SimpleNamespace(prefix=prefix, subfolder=subfolder)
+
+
+class _FakeTransformer:
+    def __init__(self) -> None:
+        self._params = {
+            "img_in.qweight": 1,
+            "img_in.qweight_type": 1,
+            "txt_in.qweight": 1,
+            "txt_in.qweight_type": 1,
+            "proj_out.qweight": 1,
+            "proj_out.qweight_type": 1,
+            "time_text_embed.timestep_embedder.linear_1.qweight": 1,
+            "time_text_embed.timestep_embedder.linear_1.qweight_type": 1,
+            "time_text_embed.timestep_embedder.linear_2.qweight": 1,
+            "time_text_embed.timestep_embedder.linear_2.qweight_type": 1,
+            "transformer_blocks.0.img_mod.1.qweight": 1,
+            "transformer_blocks.0.img_mod.1.qweight_type": 1,
+            "transformer_blocks.0.txt_mod.1.qweight": 1,
+            "transformer_blocks.0.txt_mod.1.qweight_type": 1,
+            "transformer_blocks.0.attn.to_qkv.qweight": 1,
+            "transformer_blocks.0.attn.to_qkv.qweight_type": 1,
+            "transformer_blocks.0.attn.add_kv_proj.qweight": 1,
+            "transformer_blocks.0.attn.add_kv_proj.qweight_type": 1,
+            "transformer_blocks.0.attn.to_out.weight": 1,
+            "transformer_blocks.0.img_mlp.net.0.proj.qweight": 1,
+            "transformer_blocks.0.img_mlp.net.0.proj.qweight_type": 1,
+        }
+
+    def named_parameters(self):
+        return list(self._params.items())
+
+    def named_buffers(self):
+        return []
+
+    def get_submodule(self, _prefix: str):
+        return self
+
+
+def test_qwen_adapter_selected_for_qwen_image_family():
+    adapter = get_gguf_adapter(
+        "dummy.gguf",
+        _FakeTransformer(),
+        _make_source(),
+        _make_od_config(),
+    )
+    assert isinstance(adapter, QwenImageGGUFAdapter)
+
+
+def test_qwen_adapter_matches_multiple_pipeline_variants():
+    for model_class_name in (
+        "QwenImagePipeline",
+        "QwenImageEditPipeline",
+        "QwenImageEditPlusPipeline",
+        "QwenImageLayeredPipeline",
+    ):
+        assert QwenImageGGUFAdapter.is_compatible(
+            _make_od_config(model_class_name=model_class_name),
+            _FakeTransformer(),
+            _make_source(),
+        )
+
+
+def test_qwen_adapter_preserves_split_projection_names(monkeypatch: pytest.MonkeyPatch):
+    import vllm_omni.diffusion.model_loader.gguf_adapters.qwen_image as qwen_image_module
+
+    monkeypatch.setattr(
+        qwen_image_module,
+        "gguf_quant_weights_iterator",
+        lambda _path: iter(
+            [
+                ("transformer_blocks.0.attn.to_q.qweight_type", 1),
+                ("transformer_blocks.0.attn.to_q.qweight", 2),
+                ("transformer_blocks.0.attn.to_k.qweight_type", 3),
+                ("transformer_blocks.0.attn.to_k.qweight", 4),
+                ("transformer_blocks.0.attn.to_out.0.qweight_type", 5),
+                ("transformer_blocks.0.attn.to_out.0.qweight", 6),
+            ]
+        ),
+    )
+
+    adapter = QwenImageGGUFAdapter(
+        "dummy.gguf",
+        _FakeTransformer(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    weights = list(adapter.weights_iterator())
+
+    assert ("transformer_blocks.0.attn.to_q.qweight_type", 1) in weights
+    assert ("transformer_blocks.0.attn.to_q.qweight", 2) in weights
+    assert ("transformer_blocks.0.attn.to_k.qweight_type", 3) in weights
+    assert ("transformer_blocks.0.attn.to_k.qweight", 4) in weights
+    assert ("transformer_blocks.0.attn.to_out.0.qweight_type", 5) in weights
+    assert ("transformer_blocks.0.attn.to_out.0.qweight", 6) in weights
+
+
+def test_qwen_adapter_keeps_top_level_quantized_weights(monkeypatch: pytest.MonkeyPatch):
+    import vllm_omni.diffusion.model_loader.gguf_adapters.qwen_image as qwen_image_module
+
+    monkeypatch.setattr(
+        qwen_image_module,
+        "gguf_quant_weights_iterator",
+        lambda _path: iter(
+            [
+                ("img_in.qweight_type", 1),
+                ("img_in.qweight", 2),
+                ("time_text_embed.timestep_embedder.linear_1.qweight_type", 3),
+                ("time_text_embed.timestep_embedder.linear_1.qweight", 4),
+            ]
+        ),
+    )
+
+    adapter = QwenImageGGUFAdapter(
+        "dummy.gguf",
+        _FakeTransformer(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    weights = list(adapter.weights_iterator())
+
+    assert ("img_in.qweight_type", 1) in weights
+    assert ("img_in.qweight", 2) in weights
+    assert ("time_text_embed.timestep_embedder.linear_1.qweight_type", 3) in weights
+    assert ("time_text_embed.timestep_embedder.linear_1.qweight", 4) in weights
+
+
+def test_qwen_adapter_keeps_modulation_quantized_weights(monkeypatch: pytest.MonkeyPatch):
+    import vllm_omni.diffusion.model_loader.gguf_adapters.qwen_image as qwen_image_module
+
+    monkeypatch.setattr(
+        qwen_image_module,
+        "gguf_quant_weights_iterator",
+        lambda _path: iter(
+            [
+                ("transformer_blocks.0.img_mod.1.qweight_type", 1),
+                ("transformer_blocks.0.img_mod.1.qweight", 2),
+                ("transformer_blocks.0.txt_mod.1.qweight_type", 3),
+                ("transformer_blocks.0.txt_mod.1.qweight", 4),
+            ]
+        ),
+    )
+
+    adapter = QwenImageGGUFAdapter(
+        "dummy.gguf",
+        _FakeTransformer(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    weights = list(adapter.weights_iterator())
+
+    assert ("transformer_blocks.0.img_mod.1.qweight_type", 1) in weights
+    assert ("transformer_blocks.0.img_mod.1.qweight", 2) in weights
+    assert ("transformer_blocks.0.txt_mod.1.qweight_type", 3) in weights
+    assert ("transformer_blocks.0.txt_mod.1.qweight", 4) in weights

tests/diffusion/model_loader/gguf_adapters/test_z_image_gguf_adapter.py

@@ -0,0 +1,83 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for the Z-Image GGUF adapter."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.model_loader.gguf_adapters import get_gguf_adapter
+from vllm_omni.diffusion.model_loader.gguf_adapters.z_image import (
+    ZImageGGUFAdapter,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+def _make_od_config(
+    model_class_name: str = "ZImagePipeline",
+    model_type: str = "z_image",
+):
+    return SimpleNamespace(
+        model_class_name=model_class_name,
+        tf_model_config={"model_type": model_type},
+    )
+
+
+def _make_source(prefix: str = "", subfolder: str = "transformer"):
+    return SimpleNamespace(prefix=prefix, subfolder=subfolder)
+
+
+def test_z_image_adapter_selected_for_z_image_family():
+    adapter = get_gguf_adapter(
+        "dummy.gguf",
+        object(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    assert isinstance(adapter, ZImageGGUFAdapter)
+
+
+def test_z_image_adapter_matches_model_type_variants():
+    for model_type in ("z_image", "zimage", "z-image"):
+        assert ZImageGGUFAdapter.is_compatible(
+            _make_od_config(model_class_name="OtherPipeline", model_type=model_type),
+            object(),
+            _make_source(),
+        )
+
+
+def test_z_image_adapter_renames_known_gguf_tensor_paths(monkeypatch: pytest.MonkeyPatch):
+    import vllm_omni.diffusion.model_loader.gguf_adapters.z_image as z_image_module
+
+    monkeypatch.setattr(
+        z_image_module,
+        "gguf_quant_weights_iterator",
+        lambda _path: iter(
+            [
+                ("model.diffusion_model.final_layer.qweight", torch.ones((1, 1))),
+                ("model.diffusion_model.x_embedder.qweight_type", torch.tensor(1)),
+                ("transformer_blocks.0.attention.out.weight", torch.full((1, 1), 2.0)),
+                ("transformer_blocks.0.attention.qkv.qweight", torch.full((1, 1), 3.0)),
+            ]
+        ),
+    )
+
+    adapter = ZImageGGUFAdapter(
+        "dummy.gguf",
+        object(),
+        _make_source(),
+        _make_od_config(),
+    )
+
+    weights = list(adapter.weights_iterator())
+    names = [name for name, _ in weights]
+
+    assert "all_final_layer.2-1.qweight" in names
+    assert "all_x_embedder.2-1.qweight_type" in names
+    assert "transformer_blocks.0.attention.to_out.0.weight" in names
+    assert "transformer_blocks.0.attention.to_qkv.qweight" in names

tests/diffusion/model_loader/test_diffusers_loader.py

@@ -16,6 +16,7 @@ from vllm.config.load import LoadConfig
 from vllm_omni.diffusion.config import get_current_diffusion_config, get_current_diffusion_config_or_none
 from vllm_omni.diffusion.data import OmniDiffusionConfig
 from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
+from vllm_omni.diffusion.model_loader.gguf_adapters import get_gguf_adapter
 from vllm_omni.diffusion.models.helios import HeliosPipeline
 from vllm_omni.diffusion.registry import initialize_model
 
@@ -110,6 +111,27 @@ def test_empty_source_prefix_keeps_full_model_strict_check():
         loader.load_weights(model)
 
 
+def test_qwen_model_class_selects_qwen_gguf_adapter():
+    od_config = type(
+        "Config",
+        (),
+        {
+            "model_class_name": "QwenImagePipeline",
+            "tf_model_config": {"model_type": "qwen_image"},
+        },
+    )()
+    source = DiffusersPipelineLoader.ComponentSource(
+        model_or_path="dummy",
+        subfolder="transformer",
+        revision=None,
+        prefix="transformer.",
+    )
+
+    adapter = get_gguf_adapter("dummy.gguf", object(), source, od_config)
+
+    assert adapter.__class__.__name__ == "QwenImageGGUFAdapter"
+
+
 class _ConfigAwareModel(nn.Module):
     def __init__(self, *, od_config):
         super().__init__()
@@ -164,6 +186,7 @@ def test_load_model_custom_pipeline_sets_current_diffusion_config(monkeypatch):
     loader = DiffusersPipelineLoader(LoadConfig(), od_config)
     loader.load_weights = lambda model: None  # type: ignore[assignment]
     loader._process_weights_after_loading = lambda model, target_device: None  # type: ignore[assignment]
+    loader._is_gguf_quantization = lambda: False  # type: ignore[assignment]
 
     monkeypatch.setattr(loader_mod, "resolve_obj_by_qualname", lambda _name: _ConfigAwareModel)
     monkeypatch.setattr(loader_mod.torch, "device", lambda _name: _DeviceContext("cpu"))

tests/diffusion/model_loader/test_diffusers_loader_gguf.py

@@ -0,0 +1,259 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for GGUF-specific DiffusersPipelineLoader behavior."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+import torch.nn as nn
+
+import vllm_omni.diffusion.model_loader.diffusers_loader as loader_module
+from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+class _DummyModel(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.transformer = nn.Linear(2, 2, bias=True)
+        self.vae = nn.Linear(2, 2, bias=False)
+        self.register_buffer("transformer_buffer", torch.ones(1))
+        self.calls: list[list[str]] = []
+        self.weights_sources = [
+            DiffusersPipelineLoader.ComponentSource(
+                model_or_path="dummy",
+                subfolder="transformer",
+                revision=None,
+                prefix="transformer.",
+                fall_back_to_pt=True,
+            ),
+            DiffusersPipelineLoader.ComponentSource(
+                model_or_path="dummy",
+                subfolder="vae",
+                revision=None,
+                prefix="vae.",
+                fall_back_to_pt=True,
+            ),
+        ]
+
+    def load_weights(self, weights):
+        loadable = dict(self.named_parameters())
+        loadable.update(dict(self.named_buffers()))
+        seen: list[str] = []
+        loaded: set[str] = set()
+        for name, tensor in weights:
+            seen.append(name)
+            if name in loadable:
+                target = loadable[name]
+                target.data.copy_(tensor.to(dtype=target.dtype))
+                loaded.add(name)
+        self.calls.append(seen)
+        return loaded
+
+
+def _make_loader() -> DiffusersPipelineLoader:
+    loader = object.__new__(DiffusersPipelineLoader)
+    loader.load_config = SimpleNamespace(
+        download_dir="cache-dir",
+        ignore_patterns=["*.tmp"],
+    )
+    loader.od_config = None
+    loader.quant_config = None
+    loader.counter_before_loading_weights = 0.0
+    loader.counter_after_loading_weights = 0.0
+    return loader
+
+
+def test_is_gguf_quantization_accepts_dict_config():
+    loader = _make_loader()
+    loader.quant_config = {
+        "method": "gguf",
+        "gguf_model": "weights.gguf",
+    }
+
+    assert loader._is_gguf_quantization() is True
+
+
+def test_is_gguf_quantization_rejects_non_gguf_dict_config():
+    loader = _make_loader()
+    loader.quant_config = {
+        "method": "fp8",
+        "gguf_model": "weights.gguf",
+    }
+
+    assert loader._is_gguf_quantization() is False
+
+
+def test_is_gguf_quantization_requires_gguf_model_for_dict_config():
+    loader = _make_loader()
+    loader.quant_config = {"method": "gguf"}
+
+    with pytest.raises(ValueError, match="gguf_model"):
+        loader._is_gguf_quantization()
+
+
+def test_is_gguf_quantization_accepts_object_config():
+    loader = _make_loader()
+    loader.quant_config = SimpleNamespace(
+        get_name=lambda: "gguf",
+        gguf_model="weights.gguf",
+    )
+
+    assert loader._is_gguf_quantization() is True
+
+
+def test_is_gguf_quantization_uses_fallback_object_without_get_name():
+    loader = _make_loader()
+    loader.quant_config = SimpleNamespace(gguf_model="weights.gguf")
+
+    assert loader._is_gguf_quantization() is True
+
+
+def test_get_model_loadable_names_collects_parameters_and_buffers():
+    loader = _make_loader()
+    model = _DummyModel()
+
+    names = loader._get_model_loadable_names(model)
+
+    assert "transformer.weight" in names
+    assert "transformer.bias" in names
+    assert "transformer_buffer" in names
+
+
+def test_resolve_gguf_model_path_returns_local_file(tmp_path):
+    loader = _make_loader()
+    gguf_file = tmp_path / "model.gguf"
+    gguf_file.write_bytes(b"gguf")
+
+    resolved = loader._resolve_gguf_model_path(str(gguf_file), revision=None)
+
+    assert resolved == str(gguf_file)
+
+
+def test_resolve_gguf_model_path_downloads_explicit_gguf_filename(monkeypatch: pytest.MonkeyPatch):
+    loader = _make_loader()
+
+    monkeypatch.setattr(loader_module.os.path, "isfile", lambda _path: False)
+    monkeypatch.setattr(
+        loader_module,
+        "hf_hub_download",
+        lambda repo_id, filename, revision, cache_dir: f"{cache_dir}/{repo_id}/{filename}",
+    )
+
+    resolved = loader._resolve_gguf_model_path("owner/repo/model-Q4_0.gguf", revision="main")
+
+    assert resolved == "cache-dir/owner/repo/model-Q4_0.gguf"
+
+
+def test_resolve_gguf_model_path_downloads_by_quant_type(monkeypatch: pytest.MonkeyPatch):
+    loader = _make_loader()
+
+    monkeypatch.setattr(loader_module.os.path, "isfile", lambda _path: False)
+    monkeypatch.setattr(
+        loader_module,
+        "download_gguf",
+        lambda repo_id, quant_type, cache_dir, revision, ignore_patterns: f"{cache_dir}/{repo_id}/{quant_type}.gguf",
+    )
+
+    resolved = loader._resolve_gguf_model_path("owner/repo:Q4_0", revision="main")
+
+    assert resolved == "cache-dir/owner/repo/Q4_0.gguf"
+
+
+def test_get_gguf_weights_iterator_prefixes_source_names(monkeypatch: pytest.MonkeyPatch):
+    loader = _make_loader()
+    loader.quant_config = SimpleNamespace(gguf_model="weights.gguf")
+    loader.od_config = SimpleNamespace(revision="main")
+    source = DiffusersPipelineLoader.ComponentSource(
+        model_or_path="dummy",
+        subfolder="transformer",
+        revision=None,
+        prefix="transformer.",
+    )
+
+    class _Adapter:
+        def weights_iterator(self):
+            yield "weight", torch.ones((2, 2))
+            yield "bias", torch.zeros(2)
+
+    monkeypatch.setattr(loader, "_resolve_gguf_model_path", lambda gguf_model, revision: "resolved.gguf")
+    monkeypatch.setattr(loader_module, "get_gguf_adapter", lambda gguf_file, model, source, od_config: _Adapter())
+
+    weights = list(loader._get_gguf_weights_iterator(source, object()))
+
+    assert weights[0][0] == "transformer.weight"
+    assert torch.equal(weights[0][1], torch.ones((2, 2)))
+    assert weights[1][0] == "transformer.bias"
+    assert torch.equal(weights[1][1], torch.zeros(2))
+
+
+def test_load_weights_with_gguf_falls_back_only_for_missing_transformer_weights(monkeypatch: pytest.MonkeyPatch):
+    loader = _make_loader()
+    model = _DummyModel()
+
+    monkeypatch.setattr(loader, "_get_weight_sources", lambda _model: tuple(model.weights_sources))
+    monkeypatch.setattr(
+        loader,
+        "_get_gguf_weights_iterator",
+        lambda source, model: iter([("transformer.weight", torch.ones((2, 2), dtype=torch.float32))]),
+    )
+
+    def _hf_iter(source):
+        if source.subfolder == "transformer":
+            return iter([("transformer.bias", torch.zeros(2, dtype=torch.float32))])
+        return iter([("vae.weight", torch.full((2, 2), 2.0, dtype=torch.float32))])
+
+    monkeypatch.setattr(loader, "_get_weights_iterator", _hf_iter)
+    monkeypatch.setattr(
+        loader,
+        "_get_expected_parameter_names",
+        lambda _model: {"transformer.weight", "transformer.bias", "vae.weight"},
+    )
+
+    loaded = loader._load_weights_with_gguf(model)
+
+    assert loaded == {"transformer.weight", "transformer.bias", "vae.weight"}
+    assert model.calls[0] == ["transformer.weight"]
+    assert model.calls[1] == ["transformer.bias"]
+    assert model.calls[2] == ["vae.weight"]
+
+
+def test_load_weights_with_gguf_skips_transformer_fallback_when_gguf_is_complete(monkeypatch: pytest.MonkeyPatch):
+    loader = _make_loader()
+    model = _DummyModel()
+
+    monkeypatch.setattr(loader, "_get_weight_sources", lambda _model: tuple(model.weights_sources))
+    monkeypatch.setattr(
+        loader,
+        "_get_gguf_weights_iterator",
+        lambda source, model: iter(
+            [
+                ("transformer.weight", torch.ones((2, 2), dtype=torch.float32)),
+                ("transformer.bias", torch.zeros(2, dtype=torch.float32)),
+            ]
+        ),
+    )
+
+    hf_calls: list[str] = []
+
+    def _hf_iter(source):
+        hf_calls.append(source.subfolder or "")
+        if source.subfolder == "transformer":
+            return iter([("transformer.bias", torch.zeros(2, dtype=torch.float32))])
+        return iter([("vae.weight", torch.full((2, 2), 2.0, dtype=torch.float32))])
+
+    monkeypatch.setattr(loader, "_get_weights_iterator", _hf_iter)
+    monkeypatch.setattr(
+        loader,
+        "_get_expected_parameter_names",
+        lambda _model: {"transformer.weight", "transformer.bias", "vae.weight"},
+    )
+
+    loaded = loader._load_weights_with_gguf(model)
+
+    assert loaded == {"transformer.weight", "transformer.bias", "vae.weight"}
+    assert hf_calls == ["vae"]

tests/diffusion/models/lingbot_video/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

tests/diffusion/models/lingbot_video/test_lingbot_video_transformer.py

@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
+
+
+def _tiny_transformer(**overrides):
+    from vllm_omni.diffusion.models.lingbot_video import LingBotVideoTransformer3DModel
+
+    config = {
+        "patch_size": (1, 1, 1),
+        "in_channels": 2,
+        "out_channels": 2,
+        "hidden_size": 16,
+        "num_attention_heads": 1,
+        "depth": 0,
+        "intermediate_size": 32,
+        "text_dim": 8,
+        "freq_dim": 8,
+        "axes_dims": (4, 4, 8),
+        "axes_lens": (32, 32, 32),
+    }
+    config.update(overrides)
+    return LingBotVideoTransformer3DModel(**config)
+
+
+def test_joint_position_ids_video_then_text_order():
+    from vllm_omni.diffusion.models.lingbot_video.lingbot_video_transformer import make_joint_position_ids
+
+    positions = make_joint_position_ids(text_len=3, grid_t=1, grid_h=2, grid_w=2, device=torch.device("cpu"))
+
+    assert positions.shape == (7, 3)
+    assert positions[:4, 0].tolist() == [4, 4, 4, 4]
+    assert positions[:4, 1:].tolist() == [[0, 0], [0, 1], [1, 0], [1, 1]]
+    assert positions[4:].tolist() == [[1, 0, 0], [2, 0, 0], [3, 0, 0]]
+
+
+def test_tiny_transformer_depth_zero_forward_shape():
+    model = _tiny_transformer()
+    hidden_states = torch.randn(1, 2, 1, 2, 2)
+    timestep = torch.tensor([300.0])
+    encoder_hidden_states = torch.randn(1, 3, 8)
+    encoder_attention_mask = torch.ones(1, 3, dtype=torch.long)
+
+    with torch.no_grad():
+        out = model(
+            hidden_states,
+            timestep,
+            encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            return_dict=False,
+        )[0]
+
+    assert out.shape == hidden_states.shape
+    assert torch.isfinite(out).all()
+
+
+def test_packed_attention_uses_sdpa_fallback_without_flash_varlen(monkeypatch):
+    from vllm_omni.diffusion.models.lingbot_video import lingbot_video_transformer as module
+
+    monkeypatch.setattr(module, "flash_attn_varlen_func_v3", None)
+    attn = module.LingBotVideoAttention(
+        hidden_size=8,
+        num_heads=2,
+        norm_eps=1e-6,
+        qkv_bias=False,
+        out_bias=False,
+    )
+    captured = {}
+
+    def fake_sdpa_forward(query, key, value, attn_metadata):
+        captured["mask"] = attn_metadata.attn_mask
+        return torch.zeros_like(query)
+
+    monkeypatch.setattr(attn.attn.sdpa_fallback, "forward", fake_sdpa_forward)
+    x = torch.randn(1, 5, 8)
+    rotary = torch.ones(1, 5, 2, dtype=torch.complex64)
+    packed_indices = {
+        "cu_seqlens_kv": torch.tensor([0, 2, 5], dtype=torch.int32),
+        "max_seqlen_in_batch_kv": 3,
+        "attention_mask": module._packed_block_attention_mask([2, 3], x.device),
+    }
+
+    out = attn(x, rotary, packed_indices=packed_indices)
+
+    assert out.shape == x.shape
+    mask = captured["mask"]
+    assert mask.shape == (1, 1, 5, 5)
+    assert mask[0, 0, :2, :2].all()
+    assert mask[0, 0, 2:, 2:].all()
+    assert not mask[0, 0, :2, 2:].any()
+    assert not mask[0, 0, 2:, :2].any()
+
+
+def test_tiny_transformer_rejects_invalid_rope_dims():
+    from vllm_omni.diffusion.models.lingbot_video import LingBotVideoTransformer3DModel
+
+    with pytest.raises(AssertionError, match="head_dim"):
+        LingBotVideoTransformer3DModel(
+            hidden_size=16,
+            num_attention_heads=1,
+            axes_dims=(4, 4, 4),
+            depth=0,
+        )
+
+
+def test_transformer_to_keeps_sensitive_modules_in_fp32():
+    model = _tiny_transformer()
+
+    model.to(dtype=torch.bfloat16)
+
+    assert model.patch_embedder.weight.dtype == torch.bfloat16
+    assert model.time_embedder.linear_1.weight.dtype == torch.float32
+    assert model.norm_out_modulation[1].weight.dtype == torch.float32

tests/diffusion/models/lingbot_video/test_pipeline_lingbot_video.py

@@ -0,0 +1,195 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+import torch
+from torch import nn
+
+from vllm_omni.diffusion.request import OmniDiffusionRequest
+from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
+
+
+def _make_pipeline():
+    from vllm_omni.diffusion.models.lingbot_video.pipeline_lingbot_video import LingBotVideoPipeline
+
+    pipeline = object.__new__(LingBotVideoPipeline)
+    nn.Module.__init__(pipeline)
+    pipeline.device = torch.device("cpu")
+    pipeline.default_negative_prompt = "default negative"
+    pipeline.od_config = SimpleNamespace(flow_shift=None)
+    return pipeline
+
+
+def _make_request_batch(prompt, **sampling_overrides):
+    sampling = OmniDiffusionSamplingParams(**sampling_overrides)
+    request = OmniDiffusionRequest(prompt=prompt, sampling_params=sampling, request_id="lingbot-test")
+    return DiffusionRequestBatch([request])
+
+
+def test_lingbot_video_pipeline_import_and_registry():
+    from vllm_omni.diffusion.models.lingbot_video import (
+        LingBotVideoPipeline,
+        LingBotVideoTransformer3DModel,
+        get_lingbot_video_post_process_func,
+    )
+    from vllm_omni.diffusion.registry import _DIFFUSION_MODELS, _DIFFUSION_POST_PROCESS_FUNCS
+
+    assert LingBotVideoPipeline is not None
+    assert LingBotVideoTransformer3DModel is not None
+    assert get_lingbot_video_post_process_func is not None
+    assert _DIFFUSION_MODELS["LingBotVideoPipeline"] == (
+        "lingbot_video",
+        "pipeline_lingbot_video",
+        "LingBotVideoPipeline",
+    )
+    assert _DIFFUSION_POST_PROCESS_FUNCS["LingBotVideoPipeline"] == "get_lingbot_video_post_process_func"
+
+
+def test_component_discovery_declarations():
+    from vllm_omni.diffusion.models.lingbot_video import LingBotVideoPipeline
+
+    assert LingBotVideoPipeline._dit_modules == ["transformer"]
+    assert LingBotVideoPipeline._encoder_modules == ["text_encoder"]
+    assert LingBotVideoPipeline._vae_modules == ["vae"]
+    assert LingBotVideoPipeline.supports_step_execution is False
+
+
+def test_extra_body_params_include_video_flow_shift_alias():
+    from vllm_omni.model_extras import get_extra_body_params
+
+    params = get_extra_body_params("LingBotVideoPipeline")
+    assert "shift" in params
+    assert "flow_shift" in params
+    assert "batch_cfg" in params
+
+
+def test_extract_prompt_string_and_mapping():
+    from vllm_omni.diffusion.models.lingbot_video.pipeline_lingbot_video import _extract_prompt
+
+    string_req = SimpleNamespace(prompt="a robot waves")
+    assert _extract_prompt(string_req) == ("a robot waves", None)
+
+    mapping_req = SimpleNamespace(prompt={"prompt": "a robot walks", "negative_prompt": "low quality"})
+    assert _extract_prompt(mapping_req) == ("a robot walks", "low quality")
+
+    with pytest.raises(TypeError, match="Unsupported LingBot prompt type"):
+        _extract_prompt(SimpleNamespace(prompt=["not", "supported"]))
+
+
+def test_check_inputs_accepts_t2v_shapes_and_rejects_invalid_values():
+    from vllm_omni.diffusion.models.lingbot_video import LingBotVideoPipeline
+
+    LingBotVideoPipeline.check_inputs(height=192, width=320, num_frames=1)
+    LingBotVideoPipeline.check_inputs(height=192, width=320, num_frames=9)
+
+    with pytest.raises(ValueError, match="num_frames"):
+        LingBotVideoPipeline.check_inputs(height=192, width=320, num_frames=8)
+    with pytest.raises(ValueError, match="height"):
+        LingBotVideoPipeline.check_inputs(height=190, width=320, num_frames=9)
+
+
+def test_postprocess_keeps_torch_by_default_and_converts_np_when_requested():
+    from vllm_omni.diffusion.models.lingbot_video import get_lingbot_video_post_process_func
+
+    frames = torch.ones(2, 4, 4, 3)
+    post = get_lingbot_video_post_process_func(SimpleNamespace())
+
+    assert post(frames, SimpleNamespace(output_type="pt")) is frames
+    array = post(frames, SimpleNamespace(output_type="np"))
+    assert isinstance(array, np.ndarray)
+    assert array.shape == (2, 4, 4, 3)
+
+
+def test_forward_resolves_t2v_sampling_and_flow_shift_alias():
+    pipeline = _make_pipeline()
+    calls = []
+
+    def fake_generate(**kwargs):
+        calls.append(kwargs)
+        return torch.zeros(9, 192, 320, 3)
+
+    pipeline._generate = fake_generate
+    req = _make_request_batch(
+        {"prompt": "a robot arm", "negative_prompt": "bad hands"},
+        height=192,
+        width=320,
+        num_frames=9,
+        num_inference_steps=2,
+        guidance_scale=3.0,
+        seed=42,
+        output_type="pt",
+        extra_args={"flow_shift": 4.0, "batch_cfg": True},
+    )
+
+    out = pipeline.forward(req)
+
+    assert torch.equal(out.output, torch.zeros(9, 192, 320, 3))
+    assert len(calls) == 1
+    call = calls[0]
+    assert call["prompt"] == "a robot arm"
+    assert call["negative_prompt"] == "bad hands"
+    assert call["height"] == 192
+    assert call["width"] == 320
+    assert call["num_frames"] == 9
+    assert call["num_inference_steps"] == 2
+    assert call["guidance_scale"] == 3.0
+    assert call["shift"] == 4.0
+    assert call["batch_cfg"] is True
+    assert call["output_type"] == "pt"
+
+
+def test_forward_prefers_shift_over_flow_shift_and_defaults_negative_prompt():
+    pipeline = _make_pipeline()
+    calls = []
+
+    def fake_generate(**kwargs):
+        calls.append(kwargs)
+        return torch.zeros(5, 192, 192, 3)
+
+    pipeline._generate = fake_generate
+    req = _make_request_batch(
+        "a robot arm",
+        height=192,
+        width=192,
+        num_frames=5,
+        num_inference_steps=2,
+        guidance_scale=1.0,
+        seed=42,
+        extra_args={"shift": 5.0, "flow_shift": 4.0},
+    )
+
+    pipeline.forward(req)
+
+    assert calls[0]["shift"] == 5.0
+    assert calls[0]["negative_prompt"] == "default negative"
+
+
+def test_forward_rejects_multi_request_batches():
+    pipeline = _make_pipeline()
+    first = OmniDiffusionRequest(
+        prompt="first",
+        sampling_params=OmniDiffusionSamplingParams(seed=1),
+        request_id="lingbot-first",
+    )
+    second = OmniDiffusionRequest(
+        prompt="second",
+        sampling_params=OmniDiffusionSamplingParams(seed=2),
+        request_id="lingbot-second",
+    )
+
+    with pytest.raises(ValueError, match="only supports one request"):
+        pipeline.forward(DiffusionRequestBatch([first, second]))
+
+
+def test_load_weights_rejects_external_weight_stream():
+    pipeline = _make_pipeline()
+
+    assert pipeline.load_weights([]) == set()
+    with pytest.raises(RuntimeError, match="components are loaded directly"):
+        pipeline.load_weights([("transformer.weight", torch.zeros(1))])

tests/diffusion/quantization/test_fp8_config.py

@@ -154,6 +154,21 @@ def test_component_config_with_default():
     assert config.resolve("vae.encoder") is None
 
 
+def test_gguf_config():
+    from vllm_omni.quantization import build_quant_config
+    from vllm_omni.quantization.gguf_config import DiffusionGGUFConfig
+
+    config = build_quant_config(
+        {
+            "method": "gguf",
+            "gguf_model": "path/to/model.gguf",
+        }
+    )
+    assert isinstance(config, DiffusionGGUFConfig)
+    assert config.gguf_model == "path/to/model.gguf"
+    assert config.get_name() == "gguf"
+
+
 def test_integration_string():
     from vllm_omni.diffusion.data import OmniDiffusionConfig
 
@@ -221,7 +236,7 @@ def test_transformer_config_auto_detects_modelopt_fp8():
 def test_supported_methods_includes_vllm():
     from vllm_omni.quantization import SUPPORTED_QUANTIZATION_METHODS
 
-    for method in ["fp8", "awq", "gptq", "bitsandbytes", "modelopt"]:
+    for method in ["fp8", "gguf", "awq", "gptq", "bitsandbytes", "modelopt"]:
         assert method in SUPPORTED_QUANTIZATION_METHODS, f"{method} missing"
 
 

tests/diffusion/quantization/test_gguf_config.py

@@ -0,0 +1,128 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for diffusion GGUF quantization config and linear method."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
+
+from vllm_omni.quantization import build_quant_config
+from vllm_omni.quantization.gguf_config import (
+    UNQUANTIZED_TYPES,
+    DiffusionGGUFConfig,
+    DiffusionGGUFLinearMethod,
+    dequant_gemm_gguf,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+def test_gguf_config_creation_and_delegation():
+    config = DiffusionGGUFConfig(
+        gguf_model="weights.gguf",
+        unquantized_modules=["proj_out"],
+    )
+
+    assert config.gguf_model == "weights.gguf"
+    assert config.unquantized_modules == ["proj_out"]
+    assert config.get_name() == "gguf"
+
+
+def test_build_quant_config_builds_gguf_config():
+    config = build_quant_config(
+        "gguf",
+        gguf_model="repo/model-Q4_0.gguf",
+        unquantized_modules=["norm_out"],
+    )
+
+    assert isinstance(config, DiffusionGGUFConfig)
+    assert config.gguf_model == "repo/model-Q4_0.gguf"
+    assert config.unquantized_modules == ["norm_out"]
+
+
+def test_gguf_vllm_config_returns_diffusion_linear_method_for_linear_layers():
+    linear = object.__new__(LinearBase)
+    method = DiffusionGGUFConfig(unquantized_modules=[]).get_quant_method(linear, "transformer.img_in")
+
+    assert isinstance(method, DiffusionGGUFLinearMethod)
+
+
+def test_gguf_vllm_config_respects_unquantized_modules():
+    linear = object.__new__(LinearBase)
+    method = DiffusionGGUFConfig(unquantized_modules=["proj_out"]).get_quant_method(linear, "transformer.proj_out")
+
+    assert isinstance(method, UnquantizedLinearMethod)
+
+
+def test_gguf_vllm_config_returns_none_for_non_linear_layers():
+    method = DiffusionGGUFConfig(unquantized_modules=[]).get_quant_method(torch.nn.LayerNorm(4), "norm")
+
+    assert method is None
+
+
+def test_dequant_gemm_gguf_uses_plain_matmul_for_unquantized_types():
+    qweight_type = next(iter(UNQUANTIZED_TYPES))
+    x = torch.tensor([[1.0, 2.0]], dtype=torch.float32)
+    qweight = torch.tensor([[3.0, 4.0], [5.0, 6.0]], dtype=torch.float32)
+
+    out = dequant_gemm_gguf(x, qweight, qweight_type)
+
+    assert torch.allclose(out, x @ qweight.T)
+
+
+def test_diffusion_gguf_linear_method_applies_bias_on_unquantized_weight():
+    qweight_type = next(iter(UNQUANTIZED_TYPES))
+    method = DiffusionGGUFLinearMethod(DiffusionGGUFConfig(unquantized_modules=[]))
+    layer = SimpleNamespace(
+        qweight=torch.tensor([[1.0, 2.0], [3.0, 4.0]], dtype=torch.float32),
+        qweight_type=SimpleNamespace(weight_type=qweight_type),
+    )
+    x = torch.tensor([[[1.0, 0.5]]], dtype=torch.float32)
+    bias = torch.tensor([0.25, -0.5], dtype=torch.float32)
+
+    out = method.apply(layer, x, bias)
+
+    expected = x @ layer.qweight.T + bias
+    assert torch.allclose(out, expected)
+
+
+def test_diffusion_gguf_linear_method_concatenates_sharded_outputs():
+    qweight_type = next(iter(UNQUANTIZED_TYPES))
+    method = DiffusionGGUFLinearMethod(DiffusionGGUFConfig(unquantized_modules=[]))
+    qweight = torch.nn.Parameter(
+        torch.tensor(
+            [
+                [1.0, 0.0],
+                [0.0, 1.0],
+                [2.0, 0.0],
+                [0.0, 2.0],
+            ],
+            dtype=torch.float32,
+        ),
+        requires_grad=False,
+    )
+    qweight.shard_id = ["left", "right"]
+    qweight.shard_offset_map = {
+        "left": (0, 2, 2),
+        "right": (2, 4, 2),
+    }
+    layer = SimpleNamespace(
+        qweight=qweight,
+        qweight_type=SimpleNamespace(
+            shard_weight_type={
+                "left": qweight_type,
+                "right": qweight_type,
+            }
+        ),
+    )
+    x = torch.tensor([[1.5, 2.0]], dtype=torch.float32)
+
+    out = method.apply(layer, x)
+
+    expected_q = x @ qweight[:2].T
+    expected_k = x @ qweight[2:4].T
+    assert torch.allclose(out, torch.cat([expected_q, expected_k], dim=-1))

tests/diffusion/quantization/test_trajectory_similarity_tool.py

@@ -99,6 +99,7 @@ def test_candidate_model_can_point_to_offline_checkpoint_without_online_quantiza
         candidate_quantization_config_json=None,
         candidate_ignored_layers=None,
         ignored_layers=None,
+        candidate_gguf_model=None,
         candidate_load_format="default",
     )
 

tests/e2e/offline_inference/test_krea2_expansion.py

@@ -1,155 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-"""L4 functionality tests for the Krea 2 text-to-image diffusion pipeline (``Krea2Pipeline``).
-
-Runs against the public few-step distilled checkpoint ``krea/Krea-2-Turbo`` by default (override with the
-``KREA2_MODEL`` environment variable, e.g. ``krea/Krea-2-Raw`` for the Raw checkpoint or a local diffusers
-directory). They cover a basic functional smoke plus the layerwise-CPU-offload path that the pipeline declares
-via ``SupportsComponentDiscovery`` / ``_layerwise_offload_blocks_attrs``.
-"""
-
-import os
-
-import numpy as np
-import pytest
-
-from tests.helpers.mark import hardware_test
-from tests.helpers.runtime import OmniRunnerHandler
-from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-from vllm_omni.lora.request import LoRARequest
-from vllm_omni.lora.utils import stable_lora_int_id
-
-MODEL = os.environ.get("KREA2_MODEL", "krea/Krea-2-Turbo")
-# vLLM-Omni-compatible PEFT repackaging of krea/Krea-2-LoRA-darkbrush (264 modules, r=alpha=32).
-LORA = os.environ.get("KREA2_LORA", "NagaSaiAbhinay/Krea-2-vllm-darkbrush-LoRA")
-PROMPT = "a fox in the snow, photorealistic"
-# darkbrush is a trigger-word LoRA: the trained monochrome ink-wash style only manifests when
-# its trigger phrase is present, so the LoRA test prompt appends it for a strong, stable signal.
-LORA_TRIGGER = "monochrome ink wash style"
-LORA_PROMPT = f"a fox in the snow, {LORA_TRIGGER}"
-
-pytestmark = [
-    pytest.mark.diffusion,
-    pytest.mark.full_model,
-]
-
-
-def _sampling() -> OmniDiffusionSamplingParams:
-    # Small resolution + few steps to keep the L4 case light. guidance_scale resolution is checkpoint-aware inside
-    # the pipeline (distilled -> no-CFG, Raw -> CFG), so this stays agnostic to which checkpoint KREA2_MODEL points at.
-    return OmniDiffusionSamplingParams(
-        height=512,
-        width=512,
-        num_inference_steps=8,
-        guidance_scale=0.0,
-        seed=42,
-    )
-
-
-@hardware_test(res={"cuda": "H100"})
-@pytest.mark.parametrize("omni_runner", [(MODEL, None)], indirect=True)
-def test_krea2_text_to_image_001(omni_runner_handler: OmniRunnerHandler) -> None:
-    """Basic functional smoke: a single prompt produces a decoded image."""
-    omni_runner_handler.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})
-
-
-@hardware_test(res={"cuda": "H100"})
-@pytest.mark.parametrize(
-    "omni_runner",
-    [(MODEL, None, {"enable_layerwise_offload": True})],
-    indirect=True,
-)
-def test_krea2_layerwise_offload(omni_runner_handler: OmniRunnerHandler) -> None:
-    """Exercise layerwise CPU offload on the DiT (SupportsComponentDiscovery + _layerwise_offload_blocks_attrs)."""
-    omni_runner_handler.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})
-
-
-@hardware_test(res={"cuda": "H100"})
-@pytest.mark.cache
-@pytest.mark.parametrize("omni_runner", [(MODEL, None, {"cache_backend": "cache_dit"})], indirect=True)
-def test_krea2_cache_dit(omni_runner_handler: OmniRunnerHandler) -> None:
-    """Exercise Cache-DiT on Krea 2 via the custom Krea2Pipeline enabler (``enable_cache_for_krea2``).
-
-    Validates the docs' Cache-DiT support claim for Krea 2. ``has_separate_cfg`` is checkpoint-aware
-    (False for the distilled Turbo no-CFG path, True for the Raw CFG path); the default Turbo checkpoint
-    exercises the no-CFG branch.
-    """
-    omni_runner_handler.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})
-
-
-@hardware_test(res={"cuda": "H100"}, num_cards=2)
-@pytest.mark.parametrize(
-    "omni_runner",
-    # Standalone HSDP: use_hsdp=True with an explicit hsdp_shard_size shards the DiT weights
-    # across 2 ranks (replicate_size defaults to 1 -> world_size = 1 * 2 = 2). HSDP must keep
-    # tensor_parallel_size/data_parallel_size at 1 (see vllm_omni/diffusion/data.py __post_init__).
-    # Also validated locally at 8 GPUs (hsdp_shard_size=8); CI shard provisions 2.
-    [(MODEL, None, {"use_hsdp": True, "hsdp_shard_size": 2})],
-    indirect=True,
-)
-def test_krea2_hsdp(omni_runner_handler: OmniRunnerHandler) -> None:
-    """Exercise Hybrid Sharded Data Parallel (HSDP) weight sharding on the DiT across 2 GPUs.
-
-    Validates the docs' HSDP-support claim for Krea 2. Scheduled on the multi-GPU (distributed_cuda)
-    shard by ``num_cards=2``; skipped automatically on boxes with fewer than 2 CUDA devices.
-    """
-    omni_runner_handler.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})
-
-
-@hardware_test(res={"cuda": "H100"}, num_cards=2)
-@pytest.mark.parametrize(
-    "omni_runner",
-    # VAE patch-parallel decode: partition the VAE decode across the TP world. Mirrors the online
-    # "tp_vae_patch" case (tensor_parallel_size=2 + vae_patch_parallel_size=2) from the wan2.2/z-image
-    # expansion suites, so vae_patch_parallel_size must equal the world size established by TP.
-    # Also validated locally at 8 GPUs (tp=8, vae_patch=8); CI shard provisions 2.
-    [(MODEL, None, {"tensor_parallel_size": 2, "vae_patch_parallel_size": 2})],
-    indirect=True,
-)
-def test_krea2_vae_patch_parallel(omni_runner_handler: OmniRunnerHandler) -> None:
-    """Exercise VAE patch-parallel (decode) across 2 GPUs.
-
-    Validates the docs' VAE-patch-parallel (decode) support claim for Krea 2. Scheduled on the
-    multi-GPU (distributed_cuda) shard by ``num_cards=2``; skipped automatically on boxes with
-    fewer than 2 CUDA devices.
-    """
-    omni_runner_handler.send_diffusion_request({"model": MODEL, "prompt": PROMPT, "sampling_params": _sampling()})
-
-
-def _generate(handler: OmniRunnerHandler, lora_request: LoRARequest | None, lora_scale: float = 1.0) -> np.ndarray:
-    sp = _sampling()
-    sp.lora_request = lora_request
-    sp.lora_scale = lora_scale
-    resp = handler.send_diffusion_request({"model": MODEL, "prompt": LORA_PROMPT, "sampling_params": sp})
-    return np.asarray(resp.images[0].convert("RGB"), dtype=np.int16)
-
-
-@hardware_test(res={"cuda": "H100"})
-@pytest.mark.parametrize("omni_runner", [(MODEL, None)], indirect=True)
-def test_krea2_lora(omni_runner_handler: OmniRunnerHandler) -> None:
-    """Validate diffusion LoRA on Krea 2: visible effect, scale sensitivity, clean deactivation.
-
-    Uses ``LORA`` (a vLLM-Omni-compatible PEFT repackaging of ``krea/Krea-2-LoRA-darkbrush``,
-    264 modules matched via the ReplicatedLinear projections) passed per-request through
-    ``OmniDiffusionSamplingParams.lora_request`` / ``lora_scale`` — the same fields the
-    diffusion LoRA manager reads to activate/deactivate the adapter.
-    """
-    lora_request = LoRARequest(lora_name="darkbrush", lora_int_id=stable_lora_int_id(LORA), lora_path=LORA)
-
-    baseline = _generate(omni_runner_handler, None)
-    img_1x = _generate(omni_runner_handler, lora_request, lora_scale=1.0)
-    img_2x = _generate(omni_runner_handler, lora_request, lora_scale=2.0)
-    restored = _generate(omni_runner_handler, None)
-
-    diff_1x = np.abs(baseline - img_1x).mean()
-    diff_2x = np.abs(baseline - img_2x).mean()
-    diff_restored = np.abs(baseline - restored).mean()
-
-    # (a) Adapter has a visible effect at both scales.
-    assert diff_1x > 0.5, f"LoRA scale=1.0 had no visible effect: diff={diff_1x}"
-    assert diff_2x > 0.5, f"LoRA scale=2.0 had no visible effect: diff={diff_2x}"
-    # (b) Scale changes the output (stronger adapter -> larger perturbation).
-    assert not np.isclose(diff_1x, diff_2x, atol=1.0), f"LoRA scale had no effect: 1x={diff_1x:.2f}, 2x={diff_2x:.2f}"
-    # (c) Passing lora_request=None cleanly deactivates: same seed -> back to baseline (modulo fp drift).
-    assert diff_restored < 5.0, f"LoRA did not deactivate cleanly: diff_restored={diff_restored:.2f}"

tests/e2e/offline_inference/test_qwen3_tts_customvoice.py

@@ -34,10 +34,12 @@ def get_cuda_graph_config():
             "stages": {
                 0: {
                     "max_num_seqs": 1,
+                    "gpu_memory_utilization": 0.2,
                     "enforce_eager": True,
                     "async_scheduling": False,
                 },
                 1: {
+                    "gpu_memory_utilization": 0.2,
                     "enforce_eager": True,
                     "async_scheduling": False,
                 },

tests/e2e/online_serving/test_krea2_expansion.py

@@ -1,152 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-"""L4 online-serving tests for the Krea 2 text-to-image diffusion pipeline.
-
-Exercises the OpenAI-compatible ``/v1/images/generations`` route (served via
-``vllm serve <model> --omni --enforce-eager``) rather than the offline engine path
-covered by ``tests/e2e/offline_inference/test_krea2_expansion.py``.
-
-Runs against the public few-step distilled checkpoint ``krea/Krea-2-Turbo`` by default
-(override with the ``KREA2_MODEL`` environment variable, e.g. ``krea/Krea-2-Raw`` for the
-Raw checkpoint or a local diffusers directory). Krea 2 shares the Qwen-Image VAE and a
-similar single-stream MMDiT, so this test mirrors the structure/fixtures/markers of
-``test_qwen_image_expansion.py``.
-
-Coverage mirrors the offline suite: a basic functional smoke plus the layerwise-CPU-offload
-path that the pipeline declares via ``SupportsComponentDiscovery`` /
-``_layerwise_offload_blocks_attrs``. Sizing is kept light (512x512, few steps) so the case
-fits single-GPU H100 CI rather than the full 2048x2048 distilled resolution.
-"""
-
-import base64
-import os
-from io import BytesIO
-
-import numpy as np
-import pytest
-import requests
-from PIL import Image
-
-from tests.helpers.mark import hardware_marks
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler, dummy_messages_from_mix_data
-
-pytestmark = [pytest.mark.diffusion, pytest.mark.full_model]
-
-MODEL = os.environ.get("KREA2_MODEL", "krea/Krea-2-Turbo")
-# vLLM-Omni-compatible PEFT repackaging of krea/Krea-2-LoRA-darkbrush (264 modules, r=alpha=32).
-LORA = os.environ.get("KREA2_LORA", "NagaSaiAbhinay/Krea-2-vllm-darkbrush-LoRA")
-PROMPT = "a fox in the snow, photorealistic"
-# darkbrush is a trigger-word LoRA; include its trigger phrase so the LoRA case has a strong signal.
-LORA_PROMPT = "a fox in the snow, monochrome ink wash style"
-
-SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
-
-
-def _get_diffusion_feature_cases(model: str):
-    """Return single-card L4 online-serving cases for Krea 2.
-
-    Mirrors the offline suite: a basic functional smoke plus layerwise CPU offload.
-    """
-    return [
-        # Basic functional smoke (single-card).
-        pytest.param(
-            OmniServerParams(
-                model=model,
-                server_args=["--enforce-eager"],
-            ),
-            id="single_card_basic",
-            marks=SINGLE_CARD_FEATURE_MARKS,
-        ),
-        # Layerwise CPU offload (single-card): SupportsComponentDiscovery +
-        # _layerwise_offload_blocks_attrs.
-        pytest.param(
-            OmniServerParams(
-                model=model,
-                server_args=["--enforce-eager", "--enable-layerwise-offload"],
-            ),
-            id="single_card_layerwise_offload",
-            marks=SINGLE_CARD_FEATURE_MARKS,
-        ),
-    ]
-
-
-@pytest.mark.parametrize(
-    "omni_server",
-    _get_diffusion_feature_cases(MODEL),
-    indirect=True,
-)
-def test_krea2(omni_server: OmniServer, openai_client: OpenAIClientHandler):
-    """Serve Krea 2 and validate the OpenAI-compatible image-generation route.
-
-    guidance_scale resolution is checkpoint-aware inside the pipeline (distilled -> no-CFG,
-    Raw -> CFG), so guidance_scale=0.0 stays agnostic to which checkpoint KREA2_MODEL points at.
-    Validation is delegated to assert_diffusion_response, which checks output dimensions and
-    basic correctness.
-    """
-    messages = dummy_messages_from_mix_data(content_text=PROMPT)
-    request_config = {
-        "model": omni_server.model,
-        "messages": messages,
-        "extra_body": {
-            "height": 512,
-            "width": 512,
-            "num_inference_steps": 8,
-            "guidance_scale": 0.0,
-            "seed": 42,
-        },
-    }
-    openai_client.send_diffusion_request(request_config)
-
-
-def _basic_lora_payload(lora: dict | None = None) -> dict:
-    payload = {
-        "prompt": LORA_PROMPT,
-        "n": 1,
-        "size": "512x512",
-        "num_inference_steps": 8,
-        "guidance_scale": 0.0,
-        "seed": 42,
-    }
-    if lora is not None:
-        payload["lora"] = lora
-    return payload
-
-
-def _post_image(server: OmniServer, payload: dict) -> np.ndarray:
-    url = f"http://{server.host}:{server.port}/v1/images/generations"
-    resp = requests.post(url, json=payload, headers={"Authorization": "Bearer EMPTY"}, timeout=900)
-    resp.raise_for_status()
-    b64 = resp.json()["data"][0]["b64_json"]
-    img = Image.open(BytesIO(base64.b64decode(b64)))
-    img.load()
-    return np.asarray(img.convert("RGB"), dtype=np.int16)
-
-
-@pytest.mark.parametrize(
-    "omni_server",
-    [
-        pytest.param(
-            OmniServerParams(model=MODEL, server_args=["--enforce-eager"]),
-            id="single_card_lora",
-            marks=SINGLE_CARD_FEATURE_MARKS,
-        )
-    ],
-    indirect=True,
-)
-def test_krea2_lora(omni_server: OmniServer):
-    """Serve Krea 2 and validate per-request LoRA over ``/v1/images/generations``.
-
-    Passes ``LORA`` via the top-level ``lora`` object (``{name, path, scale}``) the diffusion
-    image route parses into a ``LoRARequest`` — no server-side LoRA flag needed. Confirms the
-    adapter has a visible effect and that omitting ``lora`` cleanly restores the base output.
-    """
-    baseline = _post_image(omni_server, _basic_lora_payload())
-    with_lora = _post_image(omni_server, _basic_lora_payload({"name": "darkbrush", "path": LORA, "scale": 1.0}))
-    restored = _post_image(omni_server, _basic_lora_payload())
-
-    diff_lora = np.abs(baseline - with_lora).mean()
-    diff_restored = np.abs(baseline - restored).mean()
-
-    assert diff_lora > 0.5, f"LoRA had no visible effect over the serving path: diff={diff_lora}"
-    assert diff_restored < 5.0, f"LoRA did not deactivate cleanly: diff_restored={diff_restored:.2f}"

tests/e2e/online_serving/test_lingbot_video.py

@@ -0,0 +1,61 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Online serving smoke for ``robbyant/lingbot-video-dense-1.3b``.
+
+The LingBot dense PR supports text-to-video only, so this module exercises the
+native ``/v1/videos`` path with a small 9-frame request.
+"""
+
+import os
+
+import pytest
+
+from tests.helpers.mark import hardware_marks
+from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+MODEL = "robbyant/lingbot-video-dense-1.3b"
+PROMPT = "a robotic arm picks up a red block"
+NEGATIVE_PROMPT = "low quality, blurry, watermark, text"
+
+SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
+
+
+def _get_diffusion_feature_cases(model: str):
+    return [
+        pytest.param(
+            OmniServerParams(
+                model=model,
+                server_args=["--model-class-name", "LingBotVideoPipeline"],
+            ),
+            id="default",
+            marks=SINGLE_CARD_FEATURE_MARKS,
+        ),
+    ]
+
+
+@pytest.mark.core_model
+@pytest.mark.advanced_model
+@pytest.mark.diffusion
+@pytest.mark.parametrize("omni_server", _get_diffusion_feature_cases(MODEL), indirect=True)
+def test_text_to_video_001(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
+    request_config = {
+        "model": omni_server.model,
+        "form_data": {
+            "model": omni_server.model,
+            "prompt": PROMPT,
+            "negative_prompt": NEGATIVE_PROMPT,
+            "height": 192,
+            "width": 320,
+            "num_frames": 9,
+            "fps": 24,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.0,
+            "flow_shift": 3.0,
+            "seed": 42,
+        },
+    }
+    openai_client.send_video_diffusion_request(request_config)

tests/e2e/online_serving/test_lingbot_video_expansion.py

@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+L4 expansion coverage for ``robbyant/lingbot-video-dense-1.3b``.
+
+This is intentionally T2V-only. T2I, I2V, TI2V, MoE, and multi-GPU feature
+rows should be added in follow-up PRs when those paths are implemented.
+"""
+
+import json
+import os
+
+import pytest
+
+from tests.helpers.mark import hardware_marks
+from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+pytestmark = [pytest.mark.diffusion, pytest.mark.full_model]
+
+MODEL = "robbyant/lingbot-video-dense-1.3b"
+PROMPT = "a robotic arm picks up a red block"
+NEGATIVE_PROMPT = "low quality, blurry, watermark, text"
+
+SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
+
+
+def _get_diffusion_feature_cases(model: str):
+    return [
+        pytest.param(
+            OmniServerParams(
+                model=model,
+                server_args=["--model-class-name", "LingBotVideoPipeline"],
+            ),
+            id="default",
+            marks=SINGLE_CARD_FEATURE_MARKS,
+        ),
+    ]
+
+
+@pytest.mark.parametrize("omni_server", _get_diffusion_feature_cases(MODEL), indirect=True)
+def test_cfg_off(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
+    request_config = {
+        "model": omni_server.model,
+        "form_data": {
+            "model": omni_server.model,
+            "prompt": PROMPT,
+            "height": 192,
+            "width": 192,
+            "num_frames": 9,
+            "fps": 24,
+            "num_inference_steps": 2,
+            "guidance_scale": 1.0,
+            "flow_shift": 3.0,
+            "seed": 42,
+        },
+    }
+    openai_client.send_video_diffusion_request(request_config)
+
+
+@pytest.mark.parametrize("omni_server", _get_diffusion_feature_cases(MODEL), indirect=True)
+def test_batch_cfg_extra_params(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
+    request_config = {
+        "model": omni_server.model,
+        "form_data": {
+            "model": omni_server.model,
+            "prompt": PROMPT,
+            "negative_prompt": NEGATIVE_PROMPT,
+            "height": 192,
+            "width": 320,
+            "num_frames": 9,
+            "fps": 24,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.0,
+            "flow_shift": 3.0,
+            "seed": 42,
+            "extra_params": json.dumps({"batch_cfg": True}, separators=(",", ":")),
+        },
+    }
+    openai_client.send_video_diffusion_request(request_config)

tests/e2e/online_serving/test_qwen3_omni_expansion.py

@@ -410,44 +410,6 @@ def test_audio_in_video_002(omni_server, openai_client) -> None:
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
 
 
-@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
-def test_audio_in_video_default_loader_sampling_regression(omni_server, openai_client) -> None:
-    """
-    Regression: ``use_audio_in_video`` with default video loader sampling (no
-    ``media_io_kwargs.video.fps``) after a prior request that pinned fps.
-
-    Before the fix the second request could partial-hit mm-cache (audio cached,
-    video re-processed without audio) and fail with ``StopIteration`` in
-    ``replace_multimodal_special_tokens``.
-    """
-    video_data_url = f"data:video/mp4;base64,{generate_synthetic_video(224, 224, 128, embed_audio=True)['base64']}"
-    messages = dummy_messages_from_mix_data(
-        system_prompt=get_system_prompt(),
-        video_data_url=video_data_url,
-        content_text="Reply with one word: OK.",
-    )
-
-    base_config = {
-        "model": omni_server.model,
-        "messages": messages,
-        "stream": False,
-        "use_audio_in_video": True,
-        "modalities": ["text"],
-    }
-
-    # Prime mm-cache with explicit loader fps (historical failure mode).
-    openai_client.send_omni_request(
-        {
-            **base_config,
-            "extra_body": {"media_io_kwargs": {"video": {"fps": 1, "num_frames": 128}}},
-        }
-    )
-
-    # Default loader sampling: must succeed without explicit media_io fps.
-    openai_client.send_omni_request(dict(base_config))
-
-
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_one_word_prompt_001(omni_server, openai_client) -> None:

tests/engine/test_async_omni_engine_stage_init.py

@@ -739,7 +739,7 @@ def test_resolve_stage_configs_injects_global_diffusion_attention_when_missing(m
     monkeypatch.setattr(
         engine_mod,
         "load_and_resolve_stage_configs",
-        lambda *args, **kwargs: ("dummy-config", [stage_cfg], None),
+        lambda *args, **kwargs: ("dummy-config", [stage_cfg]),
     )
 
     _config_path, stage_configs = engine._resolve_stage_configs(
@@ -772,7 +772,7 @@ def test_resolve_stage_configs_preserves_stage_diffusion_attention(monkeypatch):
     monkeypatch.setattr(
         engine_mod,
         "load_and_resolve_stage_configs",
-        lambda *args, **kwargs: ("dummy-config", [stage_cfg], None),
+        lambda *args, **kwargs: ("dummy-config", [stage_cfg]),
     )
 
     _config_path, stage_configs = engine._resolve_stage_configs(
@@ -798,7 +798,7 @@ def test_resolve_stage_configs_does_not_inject_diffusion_attention_into_llm_stag
     monkeypatch.setattr(
         engine_mod,
         "load_and_resolve_stage_configs",
-        lambda *args, **kwargs: ("dummy-config", [stage_cfg], None),
+        lambda *args, **kwargs: ("dummy-config", [stage_cfg]),
     )
 
     _config_path, stage_configs = engine._resolve_stage_configs(
@@ -874,54 +874,3 @@ def test_build_engine_args_dict_uses_diffusion_attention_config_key():
 
     assert "attention_config" not in engine_args_dict
     assert engine_args_dict["diffusion_attention_config"].default.backend == "FLASH_ATTN"
-
-
-def test_omni_master_server_allocates_globally_unique_route_ports(monkeypatch):
-    """Regression: two stages must never draw the same ZMQ port.
-
-    ``get_open_ports_list`` only dedups within a single call, so per-route
-    allocation used to let a later stage reuse an earlier stage's port. The
-    second engine to ``bind()`` then died with ``zmq.error.ZMQError: Address
-    already in use`` (flaky multi-stage startup, e.g. Qwen3-Omni thinker/talker/
-    code2wav). ``OmniMasterServer`` now dedups every port it hands out.
-    """
-    import itertools
-
-    from vllm_omni.engine import stage_engine_startup as ses
-
-    # A colliding prefix (repeats + the master port 9000) followed by an endless
-    # fresh stream, so the only way to succeed is to redraw the collisions.
-    supply = itertools.chain([9000, 9000, 9001, 9001, 9000], itertools.count(9002))
-
-    def fake_get_open_ports_list(count):
-        return [next(supply) for _ in range(count)]
-
-    monkeypatch.setattr(ses, "get_open_ports_list", fake_get_open_ports_list)
-
-    server = ses.OmniMasterServer(
-        master_address="127.0.0.1",
-        master_port=9000,  # seed: a route must not reuse the registration port
-        stage_ids=[0, 1, 2],
-    )
-
-    ports = []
-    for sid in (0, 1, 2):
-        alloc = server.get_allocation(sid)
-        for addr in (
-            alloc.handshake_bind_address,
-            alloc.input_bind_address,
-            alloc.output_bind_address,
-        ):
-            ports.append(ses._port_from_zmq_address(addr))
-
-    assert len(ports) == len(set(ports)), f"duplicate route ports allocated: {ports}"
-    assert 9000 not in ports, "route reused the master registration port"
-
-
-def test_port_from_zmq_address_parsing():
-    from vllm_omni.engine.stage_engine_startup import _port_from_zmq_address
-
-    assert _port_from_zmq_address("tcp://127.0.0.1:34277") == 34277
-    assert _port_from_zmq_address(None) is None
-    assert _port_from_zmq_address("ipc:///tmp/sock") is None
-    assert _port_from_zmq_address("tcp://host:not-a-port") is None

tests/engine/test_strategy_lb_policy.py

@@ -1,63 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Tests for strategy-derived omni_lb_policy threading into AsyncOmniEngine.
-
-``AsyncOmniEngine._apply_strategy_lb_policy`` only touches ``self._omni_lb_policy``
-and module-level logging, so we exercise it with a lightweight stub ``self`` to
-avoid standing up a full engine (which needs GPUs/model weights).
-"""
-
-from __future__ import annotations
-
-import pytest
-
-from vllm_omni.engine.async_omni_engine import AsyncOmniEngine
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-class _Stub:
-    def __init__(self, current: str = "random") -> None:
-        self._omni_lb_policy = current
-
-
-_apply = AsyncOmniEngine._apply_strategy_lb_policy
-
-
-def test_no_derived_policy_is_noop():
-    stub = _Stub("random")
-    _apply(stub, None, {})
-    assert stub._omni_lb_policy == "random"
-
-
-def test_derived_overrides_default_random():
-    # "random" is the engine default == "unset", so the strategy value wins.
-    stub = _Stub("random")
-    _apply(stub, "round-robin", {})
-    assert stub._omni_lb_policy == "round-robin"
-
-
-def test_derived_overrides_when_flag_absent():
-    stub = _Stub("random")
-    _apply(stub, "least-queue-length", {"omni_lb_policy": None})
-    assert stub._omni_lb_policy == "least-queue-length"
-
-
-def test_explicit_matching_flag_is_fine():
-    stub = _Stub("round-robin")
-    _apply(stub, "round-robin", {"omni_lb_policy": "round-robin"})
-    assert stub._omni_lb_policy == "round-robin"
-
-
-def test_explicit_conflicting_flag_raises():
-    stub = _Stub("round-robin")
-    with pytest.raises(ValueError, match="Conflicting load-balancer policy"):
-        _apply(stub, "least-queue-length", {"omni_lb_policy": "round-robin"})
-
-
-def test_explicit_random_is_treated_as_unset():
-    # An explicit --omni-lb-policy=random is indistinguishable from the default,
-    # so a strategy value still overrides it (documented Phase-1 semantics).
-    stub = _Stub("random")
-    _apply(stub, "round-robin", {"omni_lb_policy": "random"})
-    assert stub._omni_lb_policy == "round-robin"

tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py

@@ -198,28 +198,12 @@ def _assert_realtime_smoke(result: dict) -> None:
     assert result["output_sample_rate"] > 0
 
 
-def _assert_realtime_accuracy(
+def _assert_realtime_accuracy(result: dict) -> None:
-    result: dict,
-    whisper_model_size: str = "large-v3",
-) -> None:
-    """Assert that whisper transcription of audio output matches model text.
-
-    Args:
-        result: Roundtrip result dict from ``_run_realtime_audio_roundtrip``.
-        whisper_model_size: Whisper model used to transcribe the generated audio
-                   for the accuracy check. Defaults to ``large-v3``: the default
-                   ``small`` model mishears short Chinese TTS clips (observed:
-                   北京→韦京 and a dropped leading sentence, sim=0.443), which
-                   caused spurious sim<0.8 failures under async_chunk codec
-                   variability even though audio generation was correct. large-v3
-                   transcribes these clips reliably, so a failure here now points
-                   at the model, not the ASR grader.
-    """
     final_text = (result["transcription_text"] or "").strip()
     assert final_text, "Expected non-empty transcription (model text stream)"
 
     wav_out = _wav_bytes_from_pcm16(result["output_pcm"], result["output_sample_rate"])
-    whisper_text = convert_audio_bytes_to_text(wav_out, model_size=whisper_model_size).strip()
+    whisper_text = convert_audio_bytes_to_text(wav_out).strip()
     assert whisper_text, "Whisper returned empty string for synthesized output audio"
 
     sim = cosine_similarity_text(whisper_text.lower(), final_text.lower())

tests/entrypoints/openai_api/test_serving_chat_speaker.py

@@ -87,8 +87,8 @@ def test_create_chat_completion_converts_value_error_to_error_response(mocker: M
     serving_chat.chat_template = None
     serving_chat.chat_template_content_format = "string"
     serving_chat.default_chat_template_kwargs = {}
-    serving_chat.online_renderer = mocker.MagicMock()
+    serving_chat.openai_serving_render = mocker.MagicMock()
-    serving_chat.online_renderer.validate_chat_template.return_value = None
+    serving_chat.openai_serving_render.validate_chat_template.return_value = None
     serving_chat._effective_chat_template_kwargs = mocker.MagicMock(return_value={})
     serving_chat._preprocess_chat = mocker.AsyncMock(
         side_effect=ValueError("Invalid speaker 'uncle_fu'. Supported: ethan, vivian")

tests/entrypoints/openai_api/test_serving_speech.py

@@ -39,8 +39,7 @@ from vllm_omni.entrypoints.openai.serving_speech import (
     OmniOpenAIServingSpeech,
     _create_wav_header,
 )
-from vllm_omni.entrypoints.openai.tts_adapters.base import PreparedRequest, SpeechServingContext
+from vllm_omni.entrypoints.openai.tts_adapters.base import PreparedRequest
-from vllm_omni.entrypoints.openai.tts_adapters.ming_tts import MingTTSAdapter
 from vllm_omni.model_executor.models.fish_speech.prompt_utils import (
     FISH_TEXT_ONLY_SYSTEM_PROMPT,
     build_fish_voice_clone_prompt_ids,
@@ -974,29 +973,6 @@ class TestTTSMethods:
                 )
             )
 
-    def test_upload_ming_voice_embedding_wrong_dims_rejected_without_replacing_existing(self, speech_server):
-        """Ming embedding uploads must be 192-dim before replacing stored voices."""
-        speech_server._tts_model_type = "ming_tts"
-        existing = {
-            "name": "bad_emb_voice",
-            "file_path": "/tmp/voice_samples/bad_emb_voice.safetensors",
-            "mime_type": "application/x-safetensors",
-            "embedding_source": "direct",
-            "embedding_dim": 192,
-        }
-        speech_server.uploaded_speakers = {"bad_emb_voice": existing.copy()}
-
-        with pytest.raises(ValueError, match="Ming speaker embedding must have 192 dims, got 191"):
-            asyncio.run(
-                speech_server.upload_voice_embedding(
-                    embedding_json=json.dumps([0.0] * 191),
-                    consent="consent",
-                    name="bad_emb_voice",
-                )
-            )
-
-        assert speech_server.uploaded_speakers["bad_emb_voice"] == existing
-
     def test_base_task_requires_ref_audio_or_speaker_embedding(self, speech_server):
         """Base task without ref_audio or speaker_embedding is rejected."""
         req = OpenAICreateSpeechRequest(input="Hello", task_type="Base")
@@ -1581,86 +1557,6 @@ class TestTTSMethods:
         assert params["x_vector_only_mode"] == [True]
         assert "ref_audio" not in params
 
-    def test_ming_adapter_uploaded_direct_embedding_uses_embedding_not_audio(
-        self, speech_server, mocker: MockerFixture
-    ):
-        """Ming uploaded direct embeddings are loaded into prompt construction."""
-        speech_server.uploaded_speakers = {
-            "emb_voice": {
-                "name": "emb_voice",
-                "file_path": "/tmp/voice_samples/emb_voice.safetensors",
-                "mime_type": "application/x-safetensors",
-                "embedding_source": "direct",
-                "embedding_dim": 192,
-            }
-        }
-        fake_embedding = [0.1] * 192
-        mock_get_emb = mocker.patch.object(
-            speech_server, "_get_uploaded_speaker_embedding", return_value=fake_embedding
-        )
-        mock_get_audio = mocker.patch.object(speech_server, "_get_uploaded_audio_data")
-        mock_prompt = mocker.patch.object(
-            speech_server,
-            "_build_ming_dense_prompt",
-            return_value={"additional_information": {"speaker_count": 1}},
-        )
-
-        adapter = MingTTSAdapter(SpeechServingContext(server=speech_server))
-        req = OpenAICreateSpeechRequest(input="Hello", voice="emb_voice")
-        prepared = asyncio.run(adapter.build(req, [], False))
-
-        mock_get_emb.assert_called_once_with("emb_voice")
-        mock_get_audio.assert_not_called()
-        mock_prompt.assert_called_once_with(req, ref_audio_data=None)
-        prompt_request = mock_prompt.call_args.args[0]
-        assert prompt_request.speaker_embedding == fake_embedding
-        assert len(prompt_request.speaker_embedding) == 192
-        assert prepared.prompt["additional_information"]["speaker_count"] == 1
-        assert prepared.model_type == "ming_tts"
-
-    def test_ming_adapter_uploaded_audio_path_preserves_ref_text(self, speech_server, mocker: MockerFixture):
-        """Ming uploaded audio voices still resolve audio and stored ref_text."""
-        speech_server.uploaded_speakers = {
-            "audio_voice": {
-                "name": "audio_voice",
-                "file_path": "/tmp/voice_samples/audio_voice.wav",
-                "mime_type": "audio/wav",
-                "ref_text": "Reference transcript.",
-            }
-        }
-        ref_audio_source = "data:audio/wav;base64,ZmFrZWF1ZGlv"
-        ref_audio_data = ([0.0, 0.1], 16000)
-        fake_embedding = [0.2] * 192
-        mock_get_audio = mocker.patch.object(speech_server, "_get_uploaded_audio_data", return_value=ref_audio_source)
-        mock_get_emb = mocker.patch.object(speech_server, "_get_uploaded_speaker_embedding")
-        mocker.patch.object(
-            speech_server,
-            "_resolve_ref_audio",
-            new=mocker.AsyncMock(return_value=ref_audio_data),
-        )
-        mock_extract = mocker.patch.object(
-            speech_server,
-            "_extract_ming_speaker_embeddings_from_ref_audio",
-            return_value=[fake_embedding],
-        )
-        mock_prompt = mocker.patch.object(
-            speech_server,
-            "_build_ming_dense_prompt",
-            return_value={"additional_information": {"speaker_count": 1}},
-        )
-
-        adapter = MingTTSAdapter(SpeechServingContext(server=speech_server))
-        req = OpenAICreateSpeechRequest(input="Hello", voice="audio_voice")
-        prepared = asyncio.run(adapter.build(req, [], False))
-
-        mock_get_audio.assert_called_once_with("audio_voice")
-        mock_get_emb.assert_not_called()
-        mock_extract.assert_called_once_with([ref_audio_data])
-        mock_prompt.assert_called_once_with(req, ref_audio_data=ref_audio_data)
-        assert req.ref_text == "Reference transcript."
-        assert req.speaker_embedding == fake_embedding
-        assert prepared.model_type == "ming_tts"
-
     # ── regression: full flow from issue #1603 ──
 
     def test_regression_1603_speaker_key_with_uploaded_audio_voice(self, speech_server, mocker: MockerFixture):

tests/entrypoints/test_async_omni_diffusion_config.py

@@ -343,7 +343,7 @@ def test_resolve_stage_configs_injects_additional_config_into_diffusion_stage(mo
     )
     mocker.patch(
         "vllm_omni.engine.async_omni_engine.load_and_resolve_stage_configs",
-        return_value=("dummy.yaml", [fake_llm_stage, fake_diffusion_stage], None),
+        return_value=("dummy.yaml", [fake_llm_stage, fake_diffusion_stage]),
     )
 
     engine = AsyncOmniEngine.__new__(AsyncOmniEngine)
@@ -359,15 +359,3 @@ def test_resolve_stage_configs_injects_additional_config_into_diffusion_stage(mo
 
     assert not hasattr(stage_configs[0].engine_args, "additional_config")
     assert stage_configs[1].engine_args.additional_config == {"torchair_graph_config": {"enabled": True}}
-
-
-def test_default_stage_config_includes_quantization_config():
-    """Ensure structured quantization_config survives default diffusion-stage creation."""
-    quantization_config = {
-        "method": "example_quant",
-        "weights": "weights.bin",
-    }
-
-    stage_cfg = AsyncOmniEngine._create_default_diffusion_stage_cfg({"quantization_config": quantization_config})[0]
-
-    assert stage_cfg["engine_args"]["quantization_config"] == quantization_config

tests/entrypoints/test_resolve_dreamzero_config.py

@@ -32,7 +32,7 @@ def test_dreamzero_config_sets_model_class_and_policy_config(monkeypatch):
         lambda _model: True,
     )
 
-    stage_configs, _ = load_stage_configs_from_model("GEAR-Dreams/DreamZero-DROID")
+    stage_configs = load_stage_configs_from_model("GEAR-Dreams/DreamZero-DROID")
     engine_args = stage_configs[0].engine_args
 
     assert engine_args.model_class_name == "DreamZeroPipeline"

tests/entrypoints/test_serve.py

@@ -9,7 +9,6 @@ import pytest
 from pytest_mock import MockerFixture
 
 from vllm_omni.entrypoints.cli.serve import OmniServeCommand, run_headless
-from vllm_omni.entrypoints.utils import parse_stage_overrides
 from vllm_omni.utils.tracking_parser import TrackingArgumentParser, TrackingNamespace
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -32,20 +31,6 @@ def test_serve_parser_accepts_no_async_chunk_and_marks_it_explicit() -> None:
     assert not explicit["async_chunk"]
 
 
-def test_serve_parser_accepts_strategy_config() -> None:
-    """``--strategy-config`` must parse onto the ``strategy_config`` dest and be
-    forwarded as an explicit kwarg so the engine can overlay the strategy."""
-    parser = TrackingArgumentParser()
-    subparsers = parser.add_subparsers(dest="subcommand")
-    cmd = OmniServeCommand()
-    cmd.subparser_init(subparsers)
-
-    argv = ["serve", "fake-model", "--omni", "--strategy-config", "/tmp/strategy.yaml"]
-    args = parser.parse_args(argv)
-    assert args.strategy_config == "/tmp/strategy.yaml"
-    assert args.get_explicit_kwargs_dict()["strategy_config"] == "/tmp/strategy.yaml"
-
-
 def _make_headless_args(**kwargs) -> TrackingNamespace:
     defaults = {
         "model": "fake-model",
@@ -95,97 +80,13 @@ def test_run_headless_rejects_non_multiprocess_worker_backend() -> None:
         run_headless(args)
 
 
-# ---------------------------------------------------------------------------
-# --stage-overrides parsing parity (headless vs standard path)
-# ---------------------------------------------------------------------------
-
-
-def test_parse_stage_overrides_valid_json() -> None:
-    """A valid JSON string is parsed into the nested per-stage dict."""
-    parsed = parse_stage_overrides('{"0": {"devices": "0,1"}, "1": {"devices": "2"}}')
-    assert parsed == {"0": {"devices": "0,1"}, "1": {"devices": "2"}}
-
-
-def test_parse_stage_overrides_none_and_empty_return_none() -> None:
-    """No overrides (None / empty string) resolve to ``None``."""
-    assert parse_stage_overrides(None) is None
-    assert parse_stage_overrides("") is None
-
-
-def test_parse_stage_overrides_empty_dict_returns_none() -> None:
-    """An empty dict is falsy and must resolve to ``None``, locking in parity
-    with the original standard-path ``if stage_overrides_json:`` falsy check."""
-    assert parse_stage_overrides({}) is None
-
-
-def test_parse_stage_overrides_passes_through_non_str() -> None:
-    """An already-parsed mapping is returned unchanged (identity)."""
-    overrides = {"0": {"devices": "0"}}
-    assert parse_stage_overrides(overrides) is overrides
-
-
-def test_parse_stage_overrides_invalid_json_raises() -> None:
-    """Invalid JSON raises ValueError whose message matches the standard path
-    verbatim: the ``--stage-overrides is not valid JSON:`` prefix AND the
-    ``Got: <repr>`` suffix echoing the raw input."""
-    bad = "{not valid json}"
-    with pytest.raises(ValueError) as excinfo:
-        parse_stage_overrides(bad)
-    message = str(excinfo.value)
-    assert message.startswith("--stage-overrides is not valid JSON:")
-    assert f"Got: {bad!r}" in message
-
-
-def test_run_headless_parses_and_forwards_stage_overrides(mocker: MockerFixture) -> None:
-    """Regression: the headless path must parse ``--stage-overrides`` (a JSON
-    string) and forward the parsed dict to ``load_and_resolve_stage_configs``,
-    mirroring the standard engine path. Previously it was dropped entirely,
-    silently producing a different per-stage device layout."""
-    captured: dict = {}
-
-    def _fake_resolve(*args, **kwargs):
-        captured.update(kwargs)
-        # Return a stage that does NOT match stage_id=0 so run_headless stops
-        # right after the resolver call (we only care about how it was called).
-        return ("/fake/stages.yaml", [SimpleNamespace(stage_id=99)], None)
-
-    mocker.patch(
-        "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        side_effect=_fake_resolve,
-    )
-
-    args = _make_headless_args(
-        stage_id=0,
-        strategy_config="/tmp/strategy.yaml",
-        stage_overrides='{"0": {"devices": "0,1"}, "1": {"devices": "2"}}',
-    )
-    with pytest.raises(ValueError, match="No stage config found for stage_id=0"):
-        run_headless(args)
-
-    assert captured["stage_overrides"] == {"0": {"devices": "0,1"}, "1": {"devices": "2"}}
-    assert captured["strategy_config_path"] == "/tmp/strategy.yaml"
-
-
-def test_run_headless_invalid_stage_overrides_raises(mocker: MockerFixture) -> None:
-    """Invalid ``--stage-overrides`` JSON in headless mode fails fast with the
-    shared ValueError instead of being silently ignored."""
-    mocker.patch(
-        "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        return_value=("/fake/stages.yaml", [SimpleNamespace(stage_id=0)], None),
-    )
-
-    args = _make_headless_args(stage_id=0, stage_overrides="{not valid json}")
-    with pytest.raises(ValueError, match="--stage-overrides is not valid JSON"):
-        run_headless(args)
-
-
 def test_run_headless_raises_when_stage_id_not_in_configs(mocker: MockerFixture) -> None:
     """Headless looks up its assigned stage_id in the loaded deploy YAML and
     fails fast when the launcher's --stage-id doesn't match any entry."""
     other_stage = SimpleNamespace(stage_id=99)
     mocker.patch(
         "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        return_value=("/fake/stages.yaml", [other_stage], None),
+        return_value=("/fake/stages.yaml", [other_stage]),
     )
 
     args = _make_headless_args(stage_id=0)
@@ -233,7 +134,7 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
 
     mocker.patch(
         "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        return_value=("/fake/stages.yaml", [stage_cfg], None),
+        return_value=("/fake/stages.yaml", [stage_cfg]),
     )
     mocker.patch("vllm_omni.engine.stage_init_utils.prepare_engine_environment")
     mocker.patch("vllm_omni.engine.stage_init_utils.load_omni_transfer_config_for_model", return_value=None)
@@ -309,7 +210,7 @@ def test_run_headless_llm_launches_one_manager_per_omni_dp_size_local(mocker: Mo
 
     mocker.patch(
         "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        return_value=("/fake/stages.yaml", [stage_cfg], None),
+        return_value=("/fake/stages.yaml", [stage_cfg]),
     )
     mocker.patch("vllm_omni.engine.stage_init_utils.prepare_engine_environment")
     mocker.patch("vllm_omni.engine.stage_init_utils.load_omni_transfer_config_for_model", return_value=None)
@@ -368,7 +269,7 @@ def test_run_headless_diffusion_registers_and_spawns_proc(mocker: MockerFixture)
 
     mocker.patch(
         "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        return_value=("/fake/stages.yaml", [stage_cfg], None),
+        return_value=("/fake/stages.yaml", [stage_cfg]),
     )
     mocker.patch("vllm_omni.engine.stage_init_utils.prepare_engine_environment")
     mocker.patch("vllm_omni.engine.stage_init_utils.load_omni_transfer_config_for_model", return_value=None)
@@ -448,7 +349,7 @@ def test_run_headless_diffusion_raises_on_nonzero_proc_exit(mocker: MockerFixtur
 
     mocker.patch(
         "vllm_omni.entrypoints.utils.load_and_resolve_stage_configs",
-        return_value=("/fake/stages.yaml", [stage_cfg], None),
+        return_value=("/fake/stages.yaml", [stage_cfg]),
     )
     mocker.patch("vllm_omni.engine.stage_init_utils.prepare_engine_environment")
     mocker.patch("vllm_omni.engine.stage_init_utils.load_omni_transfer_config_for_model", return_value=None)

tests/entrypoints/test_stream_finish_reason.py

@@ -144,7 +144,7 @@ def _build_serving_chat():
         engine_client=mock_engine,
         models=models,
         response_role="assistant",
-        online_renderer=mock_render,
+        openai_serving_render=mock_render,
         request_logger=None,
         chat_template=None,
         chat_template_content_format="auto",

tests/entrypoints/test_utils.py

@@ -318,7 +318,7 @@ class TestLoadAndResolveStageConfigs:
     def test_load_and_resolve_with_kwargs(self):
         """Ensure that dtype survives default stage creation."""
         kwargs = {"dtype": torch.float32}
-        config_path, stage_configs, _ = load_and_resolve_stage_configs(
+        config_path, stage_configs = load_and_resolve_stage_configs(
             model="black-forest-labs/FLUX.2-klein-4B",
             stage_configs_path=None,
             kwargs=kwargs,
@@ -349,10 +349,10 @@ class TestLoadAndResolveStageConfigs:
         ]
         load_stage_configs = mocker.patch(
             "vllm_omni.entrypoints.utils.load_stage_configs_from_model",
-            return_value=(returned_stage_configs, None),
+            return_value=returned_stage_configs,
         )
 
-        config_path, stage_configs, _ = load_and_resolve_stage_configs(
+        config_path, stage_configs = load_and_resolve_stage_configs(
             model="dummy-model",
             stage_configs_path=str(deploy_path),
             kwargs={},
@@ -363,7 +363,6 @@ class TestLoadAndResolveStageConfigs:
             base_engine_args={},
             deploy_config_path=str(deploy_path),
             stage_overrides=None,
-            strategy_config_path=None,
         )
         assert config_path == str(deploy_path)
         assert len(stage_configs) == 2

tests/helpers/runtime.py

@@ -2826,7 +2826,7 @@ def iter_omni_server(
         # be ideal to cleanup consistently everywhere.
         original_model = model_prefix + params.model
         model = original_model
-        if run_level == "core_model" and request.node.get_closest_marker("diffusion"):
+        if run_level == "core_model":
             model = resolve_tiny_model_path(model)
         port = params.port
         stage_config_path = stage_config_path_for_run_level(params.stage_config_path, run_level)
@@ -2917,7 +2917,7 @@ def iter_omni_runner(
             extra_omni_kwargs = dict(extra) if extra is not None else {}
         stage_config_path = stage_config_path_for_run_level(stage_config_path, run_level)
         model = model_prefix + model
-        if run_level == "core_model" and request.node.get_closest_marker("diffusion"):
+        if run_level == "core_model":
             model = resolve_tiny_model_path(model)
         with OmniRunner(model, seed=42, stage_configs_path=stage_config_path, **extra_omni_kwargs) as runner:
             print("OmniRunner started successfully")

tests/model_executor/models/qwen3_tts/test_code_predictor_dtype.py

@@ -18,11 +18,8 @@ import types
 
 import pytest
 import torch
-import torch.nn.functional as F
 from pytest_mock import MockerFixture
 
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
 # Direct file import to avoid vllm_omni.__init__ patch dependencies.
 _MODELS = os.path.join(
     os.path.dirname(__file__),
@@ -545,92 +542,3 @@ class TestCodePredictorWrapperConfig:
             wrapper_config=common_mod.CodePredictorWrapperConfig(use_cuda_graphs=True),
         )
         assert graph_wrapper._prefix_graphs_enabled is True
-
-
-class TestGumbelMaxSampling:
-    """Sanity tests for the shared Gumbel-max sampling helper."""
-
-    @staticmethod
-    def _sample(logits: torch.Tensor, generator=None) -> torch.Tensor:
-        common_mod = sys.modules["vllm_omni.model_executor.models.common.qwen3_code_predictor"]
-        return common_mod.CodePredictorWrapper._sample_codes_gumbel(logits, generator=generator)
-
-    def test_gumbel_max_picks_unique_high_logit(self, loaded_target_classes) -> None:
-        """If one logit is dominant, Gumbel-max must always pick it."""
-        torch.manual_seed(0)
-        bsz, vocab = 4, 64
-        logits = torch.full((bsz, vocab), -1e4)
-        logits[:, 7] = 0.0  # only finite entry
-        out = self._sample(logits)
-        assert out.shape == (bsz, 1)
-        assert torch.equal(out.squeeze(-1), torch.full((bsz,), 7, dtype=torch.long))
-
-    def test_gumbel_max_handles_single_finite_entry_rows(self, loaded_target_classes) -> None:
-        """Degenerate rows with one surviving finite entry should still return
-        a valid index instead of tripping multinomial-style asserts."""
-        bsz, vocab = 2, 8
-        logits = torch.full((bsz, vocab), float("-inf"))
-        logits[0, 3] = 0.0
-        logits[1, 5] = 0.0
-        out = self._sample(logits)
-        assert out[0].item() == 3
-        assert out[1].item() == 5
-
-    def test_gumbel_max_generator_makes_results_reproducible(self, loaded_target_classes) -> None:
-        """Same ``Generator.manual_seed`` should produce identical samples."""
-        torch.manual_seed(0)
-        logits = torch.randn(8, 32)
-        g1 = torch.Generator()
-        g1.manual_seed(123)
-        g2 = torch.Generator()
-        g2.manual_seed(123)
-        g3 = torch.Generator()
-        g3.manual_seed(456)
-        s1 = self._sample(logits, generator=g1)
-        s2 = self._sample(logits, generator=g2)
-        s3 = self._sample(logits, generator=g3)
-        assert torch.equal(s1, s2)
-        # Different seeds must (almost surely) yield a different sequence.
-        assert not torch.equal(s1, s3)
-
-    def test_gumbel_max_per_row_generators_match_individual_rows(self, loaded_target_classes) -> None:
-        """A per-row generator list should preserve row independence."""
-        torch.manual_seed(11)
-        logits = torch.randn(3, 16)
-        seeds = [101, 202, 303]
-        batched_generators = []
-        expected_rows = []
-        for row, seed in enumerate(seeds):
-            row_generator = torch.Generator()
-            row_generator.manual_seed(seed)
-            batched_generators.append(row_generator)
-
-            solo_generator = torch.Generator()
-            solo_generator.manual_seed(seed)
-            expected_rows.append(self._sample(logits[row : row + 1], generator=solo_generator))
-
-        batched = self._sample(logits, generator=batched_generators)
-        expected = torch.cat(expected_rows, dim=0)
-        assert torch.equal(batched, expected)
-
-    def test_gumbel_max_rejects_generator_length_mismatch(self, loaded_target_classes) -> None:
-        logits = torch.randn(2, 8)
-        generators = [torch.Generator()]
-        with pytest.raises(ValueError, match="Expected 2 per-row generators"):
-            self._sample(logits, generator=generators)
-
-    def test_gumbel_max_distribution_matches_softmax(self, loaded_target_classes) -> None:
-        """Empirical frequency under Gumbel-max should be close to the
-        softmax probability mass.  Loose tolerance because we draw a
-        finite number of samples."""
-        torch.manual_seed(7)
-        vocab = 16
-        logits = torch.randn(vocab) * 2.0
-        target = F.softmax(logits, dim=-1, dtype=torch.float32)
-
-        n_draws = 20000
-        batched_logits = logits.unsqueeze(0).expand(n_draws, vocab).contiguous()
-        samples = self._sample(batched_logits).squeeze(-1)
-        empirical = torch.bincount(samples, minlength=vocab).float() / n_draws
-        # L1 distance under 0.03 with 20k draws is comfortable for vocab=16.
-        assert (empirical - target).abs().sum().item() < 0.05

tests/model_executor/models/voxcpm2/test_talker_state_eviction.py

@@ -52,10 +52,6 @@ def _make_bare_talker():
     talker._cfm_buffers = None
     talker._last_audio_output_req_ids = []
     talker._batched_fsq_fusion_max_batch = 32
-    # Added in VoxCPM2TalkerForConditionalGeneration.__init__; set to
-    # default (False) to avoid AttributeError when test exercises forward
-    # path that checks this flag.
-    talker._enable_unified_decode_graph = False
     return talker
 
 
@@ -333,27 +329,23 @@ class TestDecodeBatchContract:
 
         for (expected_res, expected_meta), (actual_res, actual_meta) in zip(expected, actual, strict=True):
             torch.testing.assert_close(actual_res, expected_res)
-            torch.testing.assert_close(actual_meta.new_lm_hidden, expected_meta.new_lm_hidden)
+            torch.testing.assert_close(actual_meta["new_lm_hidden"], expected_meta["new_lm_hidden"])
         assert [state.decode_step_count for state in batch_states] == [1, 1]
 
     def test_batch_decode_matches_per_request_state_updates(self) -> None:
-        from vllm_omni.model_executor.models.voxcpm2.voxcpm2_talker import (
-            _DecodeResidualMeta,
-        )
-
         seq_talker = self._make_decode_talker()
         batch_talker = self._make_decode_talker()
 
         seq_states = [self._make_decode_state("req-0", 0.0), self._make_decode_state("req-1", 10.0)]
         batch_states = [self._make_decode_state("req-0", 0.0), self._make_decode_state("req-1", 10.0)]
         metas = [
-            _DecodeResidualMeta(new_lm_hidden=torch.tensor([[1.0, 2.0, 3.0, 4.0]])),
+            {"new_lm_hidden": torch.tensor([[1.0, 2.0, 3.0, 4.0]])},
-            _DecodeResidualMeta(new_lm_hidden=torch.tensor([[5.0, 6.0, 7.0, 8.0]])),
+            {"new_lm_hidden": torch.tensor([[5.0, 6.0, 7.0, 8.0]])},
         ]
         batch_out = torch.tensor([[0.5, 1.0, 1.5, 2.0], [2.5, 3.0, 3.5, 4.0]])
 
         for state, meta, res_out in zip(seq_states, metas, batch_out, strict=True):
-            seq_talker._finish_decode(state, meta, res_out.unsqueeze(0))
+            seq_talker._finish_decode(state, meta, res_out.unsqueeze(0), torch.device("cpu"))
 
         batch_talker._finish_decode_batch(
             [(state, False, meta) for state, meta in zip(batch_states, metas, strict=True)],
@@ -368,16 +360,12 @@ class TestDecodeBatchContract:
             torch.testing.assert_close(actual.last_audio_patch_gpu, expected.last_audio_patch_gpu)
 
     def test_batch_decode_preserves_per_request_cfm_rng_order(self) -> None:
-        from vllm_omni.model_executor.models.voxcpm2.voxcpm2_talker import (
-            _DecodeResidualMeta,
-        )
-
         talker = self._make_decode_talker()
         talker._enable_batched_cfm = False
         states = [self._make_decode_state("req-0", 0.0), self._make_decode_state("req-1", 10.0)]
         metas = [
-            _DecodeResidualMeta(new_lm_hidden=torch.tensor([[1.0, 2.0, 3.0, 4.0]])),
+            {"new_lm_hidden": torch.tensor([[1.0, 2.0, 3.0, 4.0]])},
-            _DecodeResidualMeta(new_lm_hidden=torch.tensor([[5.0, 6.0, 7.0, 8.0]])),
+            {"new_lm_hidden": torch.tensor([[5.0, 6.0, 7.0, 8.0]])},
         ]
         batch_out = torch.tensor([[0.5, 1.0, 1.5, 2.0], [2.5, 3.0, 3.5, 4.0]])
 

tests/model_executor/stage_input_processors/test_cosyvoice3_stage_input_processors.py

@@ -8,6 +8,7 @@ import torch
 
 from vllm_omni.model_executor.stage_input_processors.cosyvoice3 import (
     talker2code2wav_async_chunk,
+    text2flow,
     text2flow_full_payload,
     text2flow_token_only,
 )
@@ -48,13 +49,13 @@ def _transfer_manager(
     )
 
 
-def test_text2flow_token_only_supports_batched_source_outputs():
+def test_text2flow_supports_batched_source_outputs():
     source_outputs = [
         _source_output("req-0", [10, 11], [1, 2, 3], {"speech_token": torch.tensor([[8, 9]])}),
         _source_output("req-1", [20, 21], [4, 5], {"speech_token": torch.tensor([[6, 7]])}),
     ]
 
-    outputs = text2flow_token_only(source_outputs=source_outputs, prompt=None)
+    outputs = text2flow(source_outputs=source_outputs, prompt=None)
 
     assert len(outputs) == 2
     assert outputs[0]["prompt_token_ids"] == [1, 2, 3]
@@ -63,6 +64,16 @@ def test_text2flow_token_only_supports_batched_source_outputs():
     assert outputs[1]["additional_information"]["ids"]["prompt"] == [20, 21]
 
 
+def test_text2flow_strips_reference_speech_prefix_from_cumulative_ids():
+    source_outputs = [
+        _source_output("req-0", [10, 11], [8, 9, 1, 2, 3], {"speech_token": torch.tensor([[8, 9]])}),
+    ]
+
+    outputs = text2flow(source_outputs=source_outputs, prompt=None)
+
+    assert outputs[0]["prompt_token_ids"] == [1, 2, 3]
+
+
 def test_text2flow_token_only_strips_reference_speech_prefix_from_cumulative_ids():
     source_outputs = [
         _source_output(

tests/model_executor/stage_input_processors/test_indextts2.py

@@ -7,6 +7,7 @@ import torch
 
 from vllm_omni.model_executor.stage_input_processors.indextts2 import (
     _strip_stop_token,
+    talker2s2mel,
     talker2s2mel_full_payload,
     talker2s2mel_token_only,
 )
@@ -107,6 +108,118 @@ def test_strip_stop_token_all_stop():
     assert lat.shape[1] == 0
 
 
+# ── talker2s2mel ──
+
+
+def test_talker2s2mel_basic():
+    mel_codes = torch.tensor([10, 20, 30, STOP_MEL_TOKEN])
+    latent = torch.randn(4, LATENT_DIM)
+    s_ref = torch.randn(1, 128)
+    ref_mel = torch.randn(1, 80, 200)
+    style = torch.randn(1, 256)
+
+    outputs = talker2s2mel(
+        source_outputs=[
+            _make_talker_output(
+                mel_codes=mel_codes,
+                latent=latent,
+                s_ref=s_ref,
+                ref_mel=ref_mel,
+                style=style,
+            )
+        ]
+    )
+
+    assert len(outputs) == 1
+    info = outputs[0]["additional_information"]
+    assert info["mel_codes"].shape == (1, 3)
+    assert info["latent"].shape == (1, 3, LATENT_DIM)
+    assert info["code_lens"].tolist() == [3]
+    assert "S_ref" in info
+    assert "ref_mel" in info
+    assert "style" in info
+    assert outputs[0]["prompt_token_ids"] == [0]
+
+
+def test_talker2s2mel_skips_unfinished():
+    mel_codes = torch.tensor([10, 20])
+    latent = torch.randn(2, LATENT_DIM)
+
+    outputs = talker2s2mel(source_outputs=[_make_talker_output(mel_codes=mel_codes, latent=latent, finished=False)])
+
+    assert len(outputs) == 0
+
+
+def test_talker2s2mel_skips_empty_mel():
+    outputs = talker2s2mel(
+        source_outputs=[
+            _make_talker_output(
+                mel_codes=torch.tensor([]),
+                latent=torch.randn(0, LATENT_DIM),
+            )
+        ]
+    )
+
+    assert len(outputs) == 0
+
+
+def test_talker2s2mel_missing_optional_meta():
+    mel_codes = torch.tensor([10, 20])
+    latent = torch.randn(2, LATENT_DIM)
+
+    outputs = talker2s2mel(source_outputs=[_make_talker_output(mel_codes=mel_codes, latent=latent)])
+
+    assert len(outputs) == 1
+    info = outputs[0]["additional_information"]
+    assert "S_ref" not in info
+    assert "ref_mel" not in info
+    assert "style" not in info
+
+
+def test_talker2s2mel_batch():
+    out0 = _make_talker_output(
+        mel_codes=torch.tensor([1, 2, STOP_MEL_TOKEN]),
+        latent=torch.randn(3, LATENT_DIM),
+        s_ref=torch.randn(1, 64),
+    )
+    out1 = _make_talker_output(
+        mel_codes=torch.tensor([5, 6, 7, 8]),
+        latent=torch.randn(4, LATENT_DIM),
+        ref_mel=torch.randn(1, 80, 100),
+    )
+
+    outputs = talker2s2mel(source_outputs=[out0, out1])
+
+    assert len(outputs) == 2
+    assert outputs[0]["additional_information"]["code_lens"].tolist() == [2]
+    assert outputs[1]["additional_information"]["code_lens"].tolist() == [4]
+
+
+def test_talker2s2mel_tensors_on_cpu():
+    device = torch.device("cpu")
+    mel_codes = torch.tensor([10, 20], device=device)
+    latent = torch.randn(2, LATENT_DIM, device=device)
+    s_ref = torch.randn(1, 64, device=device)
+
+    outputs = talker2s2mel(source_outputs=[_make_talker_output(mel_codes=mel_codes, latent=latent, s_ref=s_ref)])
+
+    info = outputs[0]["additional_information"]
+    assert info["latent"].device.type == "cpu"
+    assert info["mel_codes"].device.type == "cpu"
+    assert info["S_ref"].device.type == "cpu"
+
+
+def test_talker2s2mel_no_multimodal_dict():
+    output = SimpleNamespace(
+        finished=True,
+        outputs=[SimpleNamespace(multimodal_output="not_a_dict")],
+    )
+
+    results = talker2s2mel(source_outputs=[output])
+
+    assert len(results) == 0
+
+
 # ── full-payload connector path ──
 
 

tests/model_executor/stage_input_processors/test_mimo_audio_llm2code2wav.py

@@ -11,40 +11,59 @@ import torch
 from vllm_omni.model_executor.stage_input_processors import mimo_audio as sip
 from vllm_omni.model_executor.stage_input_processors.mimo_audio import (
     MAX_CODE2WAV_TOKENS,
-    llm2code2wav_full_payload,
+    llm2code2wav,
 )
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
 
-def test_llm2code2wav_full_payload_truncates_when_flat_exceeds_max(caplog):
+def _make_source_outputs(codec_codes: torch.Tensor, request_id: str = "req-0"):
-    """Flat codec sequences longer than MAX_CODE2WAV_TOKENS must be truncated."""
+    output = SimpleNamespace(multimodal_output={"codes": {"audio": codec_codes}})
+    """Build minimal source_outputs with one talker output carrying codec codes."""
+    talker_output = SimpleNamespace(outputs=[output], request_id=request_id)
+    return [talker_output]
+
+
+def test_llm2code2wav_truncates_when_flat_exceeds_max(caplog):
+    """Flat codec sequences longer than MAX_CODE2WAV_TOKENS must be truncated, not passed through."""
+    # prepend_and_flatten_colmajor produces 36 ids per (8, 4) codec frame:
+    # pad adds one row -> (9, 4) per frame, permuted and flattened.
+    # Pick enough frames to comfortably exceed the cap.
     frames = (MAX_CODE2WAV_TOKENS // 36) + 100
     codec_codes = torch.ones(frames, 1, 8, 4, dtype=torch.long)
-    request = SimpleNamespace(request_id="req-long")
 
+    source_outputs = _make_source_outputs(codec_codes, request_id="req-long")
+
+    # Attach caplog's handler directly to the module logger so the warning is
+    # captured regardless of propagation (vllm's logger configuration can
+    # interact badly with caplog.at_level's default root-handler path).
     target_logger = logging.getLogger("vllm_omni.model_executor.stage_input_processors.mimo_audio")
     target_logger.addHandler(caplog.handler)
     prev_level = target_logger.level
     target_logger.setLevel(logging.WARNING)
     try:
-        payload = llm2code2wav_full_payload(None, {"codes.audio": codec_codes}, request)
+        prompts = llm2code2wav(source_outputs)
     finally:
         target_logger.removeHandler(caplog.handler)
         target_logger.setLevel(prev_level)
 
-    assert payload is not None
+    assert len(prompts) == 1
-    assert len(payload["codes"]["audio"]) == MAX_CODE2WAV_TOKENS
+    assert len(prompts[0]["prompt_token_ids"]) == MAX_CODE2WAV_TOKENS
+    assert any("truncating" in rec.getMessage() for rec in caplog.records), (
+        f"Expected a 'truncating' warning; captured records: {[r.getMessage() for r in caplog.records]}"
+    )
 
 
-def test_llm2code2wav_full_payload_short_sequence_unchanged():
+def test_llm2code2wav_short_sequence_unchanged():
+    """Short codec sequences are returned without truncation."""
     codec_codes = torch.ones(4, 1, 8, 4, dtype=torch.long)
-    request = SimpleNamespace(request_id="req-short")
+    source_outputs = _make_source_outputs(codec_codes, request_id="req-short")
 
-    payload = llm2code2wav_full_payload(None, {"codes.audio": codec_codes}, request)
+    prompts = llm2code2wav(source_outputs)
 
-    assert payload is not None
+    assert len(prompts) == 1
-    assert 0 < len(payload["codes"]["audio"]) <= MAX_CODE2WAV_TOKENS
+    # 4 frames + 1 pad row, flattened col-major → well below the cap
+    assert 0 < len(prompts[0]["prompt_token_ids"]) <= MAX_CODE2WAV_TOKENS
 
 
 def test_llm2code2wav_truncation_boundary_constant_matches_yaml():

tests/model_executor/stage_input_processors/test_ming_tts_async_chunk.py

@@ -1,102 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from collections import defaultdict
-from types import SimpleNamespace
-
-import pytest
-import torch
-
-from vllm_omni.model_executor.models.ming_tts.constants import LATENT_DIM, PATCH_SIZE
-from vllm_omni.model_executor.stage_input_processors.ming_tts import (
-    MING_EMIT_PATCH_COUNT_KEY,
-    llm2audio_vae_async_chunk,
-)
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def _make_transfer_manager(*, chunk_size: int = 5, initial_chunk_size: int = 2):
-    return SimpleNamespace(
-        code_prompt_token_ids=defaultdict(list),
-        put_req_chunk=defaultdict(int),
-        request_payload={},
-        connector=SimpleNamespace(
-            config={
-                "extra": {
-                    "latent_chunk_size": chunk_size,
-                    "initial_latent_chunk_size": initial_chunk_size,
-                    "latent_left_context": 0,
-                }
-            }
-        ),
-    )
-
-
-def _make_request(req_id: str = "req", *, finished: bool = False):
-    return SimpleNamespace(
-        external_req_id=req_id,
-        is_finished=lambda: finished,
-    )
-
-
-def _make_output(value: float):
-    return {
-        "ming_has_patch": torch.tensor([1], dtype=torch.bool),
-        "ming_latent_patch": torch.full((1, PATCH_SIZE, LATENT_DIM), value),
-    }
-
-
-def _append_patch(tm, req_id: str, idx: int, *, finished: bool = False):
-    return llm2audio_vae_async_chunk(
-        tm,
-        _make_output(float(idx)),
-        _make_request(req_id, finished=finished),
-        is_finished=finished,
-    )
-
-
-def test_ming_async_chunk_emits_small_initial_chunk_then_steady_chunks():
-    tm = _make_transfer_manager(chunk_size=5, initial_chunk_size=2)
-
-    assert _append_patch(tm, "req", 0) is None
-
-    first = _append_patch(tm, "req", 1)
-    assert first is not None
-    assert first.latent.shape == (2, PATCH_SIZE, LATENT_DIM)
-    assert first.kv_metadata[MING_EMIT_PATCH_COUNT_KEY] == 2
-    assert first.meta.finished.item() is False
-
-    for idx in range(2, 6):
-        assert _append_patch(tm, "req", idx) is None
-
-    second = _append_patch(tm, "req", 6)
-    assert second is not None
-    assert second.latent.shape == (5, PATCH_SIZE, LATENT_DIM)
-    assert second.kv_metadata[MING_EMIT_PATCH_COUNT_KEY] == 5
-
-
-def test_ming_async_chunk_flushes_short_final_chunk():
-    tm = _make_transfer_manager(chunk_size=5, initial_chunk_size=2)
-
-    payload = _append_patch(tm, "req", 0, finished=True)
-
-    assert payload is not None
-    assert payload.latent.shape == (1, PATCH_SIZE, LATENT_DIM)
-    assert payload.kv_metadata[MING_EMIT_PATCH_COUNT_KEY] == 1
-    assert payload.meta.finished.item() is True
-
-
-def test_ming_async_chunk_flushes_leftover_after_initial_chunk():
-    tm = _make_transfer_manager(chunk_size=5, initial_chunk_size=2)
-
-    assert _append_patch(tm, "req", 0) is None
-    assert _append_patch(tm, "req", 1) is not None
-    assert _append_patch(tm, "req", 2) is None
-
-    final = _append_patch(tm, "req", 3, finished=True)
-
-    assert final is not None
-    assert final.latent.shape == (2, PATCH_SIZE, LATENT_DIM)
-    assert final.kv_metadata[MING_EMIT_PATCH_COUNT_KEY] == 2
-    assert final.meta.finished.item() is True

tests/model_executor/stage_input_processors/test_qwen3_omni_streaming_helpers.py

@@ -163,7 +163,7 @@ def test_talker2code2wav_full_payload_filters_by_output_token_ids() -> None:
 
     assert payload is not None
     assert payload["codes"]["audio"] == [10, 20, 11, 21, 12, 22]
-    assert "code_predictor_codes" not in payload
+    assert payload["code_predictor_codes"] == payload["codes"]["audio"]
 
 
 def test_talker2code2wav_full_payload_drops_count_matched_terminal_row() -> None:
@@ -204,7 +204,7 @@ def test_talker2code2wav_full_payload_drops_rows_aligned_to_non_codec_ids() -> N
 
     assert payload is not None
     assert payload["codes"]["audio"] == [0, 0, 0]
-    assert "code_predictor_codes" not in payload
+    assert payload["code_predictor_codes"] == payload["codes"]["audio"]
 
 
 def test_talker2code2wav_full_payload_keeps_all_zero_codec_rows() -> None:
@@ -224,7 +224,7 @@ def test_talker2code2wav_full_payload_keeps_all_zero_codec_rows() -> None:
 
     assert payload is not None
     assert payload["codes"]["audio"] == [0, 7, 0, 8, 0, 9]
-    assert "code_predictor_codes" not in payload
+    assert payload["code_predictor_codes"] == payload["codes"]["audio"]
 
 
 def test_thinker2talker_full_payload_packs_complete_tensors() -> None:
@@ -247,6 +247,7 @@ def test_thinker2talker_full_payload_packs_complete_tensors() -> None:
     assert payload["ids"]["all"] == [151644, 872, 3]
     assert payload["embed"]["prefill"].device.type == "cpu"
     assert payload["hidden_states"]["output"].device.type == "cpu"
+    assert payload["next_stage_prompt_len"] > 0
     assert payload["embed"]["prefill"].shape[0] == 2
     assert payload["hidden_states"]["output"].shape[0] == 2
 

tests/model_executor/stage_input_processors/test_qwen3_tts_async_chunk.py

@@ -12,9 +12,9 @@ from vllm_omni.model_executor.stage_input_processors.chunk_size_utils import (
     max_ic_for_chunk_size,
 )
 from vllm_omni.model_executor.stage_input_processors.qwen3_tts import (
+    talker2code2wav,
     talker2code2wav_async_chunk,
     talker2code2wav_full_payload,
-    talker2code2wav_token_only,
 )
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -368,15 +368,7 @@ def test_ref_code_context_can_be_buffered_before_first_emit():
     assert rid in tm.request_payload
 
 
-def test_non_async_token_only_sizes_placeholder_for_ref_and_audio_frames():
+def test_non_async_processor_prepends_ref_code_and_sets_trim_context():
-    """``talker2code2wav_token_only`` only allocates placeholder prompt slots.
-
-    After the connector refactor, actual codec flattening (ref prepend +
-    codebook-major layout) is performed by ``talker2code2wav_full_payload`` on
-    the worker data plane.  The orchestrator hook still derives
-    ``left_context_size`` from stage-0 multimodal_output so Code2Wav can trim
-    reference frames once the connector payload arrives.
-    """
     ref_code = torch.tensor([[9, 9, 9, 9], [8, 8, 8, 8]], dtype=torch.long)
     audio_codes = torch.tensor(
         [
@@ -395,38 +387,12 @@ def test_non_async_token_only_sizes_placeholder_for_ref_and_audio_frames():
         engine_outputs=[SimpleNamespace(outputs=[output], finished=True)],
     )
 
-    prompts = talker2code2wav_token_only(stage.engine_outputs)
+    prompts = talker2code2wav(stage.engine_outputs)
 
     assert len(prompts) == 1
     prompt = prompts[0]
     assert prompt["additional_information"] == {"meta": {"left_context_size": 2}}
-    # 2 ref frames + 2 valid audio frames (zero row filtered), 4 quantizers.
+    assert prompt["prompt_token_ids"] == [
-    assert prompt["prompt_token_ids"] == [0] * (_Q * (2 + 2))
-
-
-def test_full_payload_prepends_ref_code_and_flattens_codebook_major():
-    """Worker producer is authoritative for ref prepend + codec flatten."""
-    ref_code = torch.tensor([[9, 9, 9, 9], [8, 8, 8, 8]], dtype=torch.long)
-    audio_codes = torch.tensor(
-        [
-            [0, 0, 0, 0],
-            [1, 2, 3, 4],
-            [5, 6, 7, 8],
-        ],
-        dtype=torch.long,
-    )
-    pooling_output = {
-        "codes.audio": audio_codes,
-        "codes.ref": ref_code,
-        "meta.ref_code_len": 2,
-    }
-    request = SimpleNamespace(request_id="r", output_token_ids=list(range(3)))
-
-    payload = talker2code2wav_full_payload(transfer_manager=None, pooling_output=pooling_output, request=request)
-
-    assert payload is not None
-    assert payload["meta"]["left_context_size"] == 2
-    assert payload["codes"]["audio"].tolist() == [
         9,
         8,
         1,
@@ -467,7 +433,7 @@ def test_non_async_processor_filters_out_of_range_codec_values():
         engine_outputs=[SimpleNamespace(outputs=[output], finished=True)],
     )
 
-    prompts = talker2code2wav_token_only(stage.engine_outputs)
+    prompts = talker2code2wav(stage.engine_outputs)
 
     assert len(prompts) == 1
     prompt = prompts[0]

tests/model_executor/stage_input_processors/test_voxtral_tts_async_chunk.py

@@ -8,12 +8,61 @@ import pytest
 import torch
 
 from vllm_omni.model_executor.stage_input_processors.voxtral_tts import (
+    generator2tokenizer,
     generator2tokenizer_async_chunk,
 )
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
 
+# ---- Helpers for generator2tokenizer (non-streaming) ----
+
+
+def _make_stage(engine_outputs):
+    return SimpleNamespace(engine_outputs=engine_outputs)
+
+
+def _make_output(audio_tensors):
+    """Build a single generator output with multimodal_output codes.audio."""
+    return SimpleNamespace(
+        outputs=[SimpleNamespace(multimodal_output={"codes": {"audio": audio_tensors}})],
+    )
+
+
+# ---- Tests for generator2tokenizer ----
+
+
+class TestGenerator2Tokenizer:
+    def test_single_output_single_chunk(self):
+        """Single output with one audio tensor produces correct prompt_token_ids."""
+        audio = torch.tensor([10, 20, 30, 40])
+        stage = _make_stage([_make_output([audio])])
+
+        result = generator2tokenizer(stage.engine_outputs)
+
+        assert len(result) == 1
+        assert result[0]["prompt_token_ids"] == [10, 20, 30, 40]
+        assert result[0]["multi_modal_data"] is None
+
+    def test_2d_audio_tensors_flattened(self):
+        """2D audio tensors (e.g., multi-codebook frames) are flattened."""
+        audio = torch.tensor([[1, 2, 3], [4, 5, 6]])  # (2, 3)
+        stage = _make_stage([_make_output([audio])])
+
+        result = generator2tokenizer(stage.engine_outputs)
+
+        assert result[0]["prompt_token_ids"] == [1, 2, 3, 4, 5, 6]
+
+    def test_no_outputs_yet_raises(self):
+        """Stage with engine_outputs=None raises RuntimeError."""
+        stage = _make_stage(engine_outputs=None)
+        with pytest.raises(TypeError):
+            generator2tokenizer(stage.engine_outputs)
+
+
+# ---- Helpers for generator2tokenizer_async_chunk (streaming) ----
+
+
 def _req(external_req_id: str, *, finished: bool):
     return SimpleNamespace(
         external_req_id=external_req_id,

tests/model_tests/diffusion/test_alignment.py

@@ -56,7 +56,6 @@ EXCLUDED_MODELS = [
     "NextStep11Pipeline",
     "FluxPipeline",
     "FluxDMD2Pipeline",
-    "Krea2Pipeline",
     "QwenImageDMD2Pipeline",
     "OmniGen2Pipeline",
     "HeliosPipeline",
@@ -67,6 +66,7 @@ EXCLUDED_MODELS = [
     "AudioXPipeline",
     "HunyuanVideo15Pipeline",
     "HunyuanVideo15ImageToVideoPipeline",
+    "LingBotVideoPipeline",
     "MagiHumanPipeline",
     "OmniVoicePipeline",
     "OmniVoice",

tests/platforms/test_cuda_ir_op_priority.py

@@ -1,38 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import pytest
-from vllm.config import CompilationConfig, CompilationMode, DeviceConfig, VllmConfig
-
-from vllm_omni.platforms.cuda.platform import CudaOmniPlatform
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-
-def _vllm_config(*, backend: str = "inductor", mode: CompilationMode) -> VllmConfig:
-    return VllmConfig(
-        device_config=DeviceConfig(device="cpu"),
-        compilation_config=CompilationConfig(backend=backend, mode=mode),
-    )
-
-
-@pytest.mark.parametrize(
-    "mode",
-    [CompilationMode.NONE, CompilationMode.VLLM_COMPILE, CompilationMode.STOCK_TORCH_COMPILE],
-)
-def test_cuda_default_ir_op_priority_prefers_vllm_c_when_inductor_backend(mode: CompilationMode) -> None:
-    """Regression for #4964: inductor-active configs must not switch to native-only."""
-    priority = CudaOmniPlatform.get_default_ir_op_priority(_vllm_config(mode=mode))
-    assert priority.rms_norm == ["vllm_c", "native"]
-    assert priority.fused_add_rms_norm == ["vllm_c", "native"]
-
-
-def test_cuda_default_ir_op_priority_with_oink(monkeypatch: pytest.MonkeyPatch) -> None:
-    import vllm.envs as envs
-
-    monkeypatch.setattr(envs, "VLLM_USE_OINK_OPS", True)
-    priority = CudaOmniPlatform.get_default_ir_op_priority(
-        _vllm_config(mode=CompilationMode.VLLM_COMPILE),
-    )
-    assert priority.rms_norm == ["oink", "vllm_c", "native"]
-    assert priority.fused_add_rms_norm == ["oink", "vllm_c", "native"]

tests/plugin/__init__.py

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

tests/plugin/test_gguf_diffusion.py

@@ -1,154 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""End-to-end tests for GGUF quantization on diffusion models.
-
-Validates that GGUF-quantized diffusion models generate valid images and
-use less peak GPU memory than BF16 baseline.
-
-Requires vllm-omni to be installed alongside the plugin.
-
-Usage:
-    pytest tests/diffusion/test_gguf_memory.py -v
-"""
-
-from __future__ import annotations
-
-from typing import NamedTuple
-
-import numpy as np
-import pytest
-import torch
-
-from tests.helpers.env import DeviceMemoryMonitor
-from tests.helpers.mark import hardware_test
-from tests.helpers.runtime import OmniRunner
-from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-from vllm_omni.outputs import OmniRequestOutput
-from vllm_omni.platforms import current_omni_platform
-
-
-class DiffusionGGUFTestConfig(NamedTuple):
-    artifact_prefix: str
-    gguf_model: str
-    hf_model: str
-    min_cosine_similarity: float = 0.95
-
-
-Z_IMAGE_CONFIG = DiffusionGGUFTestConfig(
-    artifact_prefix="zimage",
-    hf_model="Tongyi-MAI/Z-Image-Turbo",
-    gguf_model="unsloth/Z-Image-Turbo-GGUF:Q4_0",
-)
-
-
-FLUX_CONFIG = DiffusionGGUFTestConfig(
-    artifact_prefix="flux2_klein",
-    hf_model="black-forest-labs/FLUX.2-klein-4B",
-    gguf_model="unsloth/FLUX.2-klein-4B-GGUF:Q8_0",
-)
-
-
-def _image_cosine_similarity(hf_image, gguf_image) -> float:
-    hf_tensor = torch.as_tensor(np.array(hf_image.convert("RGB"), copy=True), dtype=torch.float32).flatten()
-    gguf_tensor = torch.as_tensor(np.array(gguf_image.convert("RGB"), copy=True), dtype=torch.float32).flatten()
-    return torch.nn.functional.cosine_similarity(hf_tensor, gguf_tensor, dim=0).item()
-
-
-def _generate_single_stage_image(
-    model: str,
-    height: int = 256,
-    width: int = 256,
-    num_inference_steps: int = 20,
-    seed: int = 42,
-    **extra_kwargs,
-) -> tuple[list, float]:
-    """Generate an image with a single-stage diffusion model.
-
-    Returns (images, peak_memory_gib).
-    """
-    omni_kwargs = dict(extra_kwargs)
-
-    memory_monitor = DeviceMemoryMonitor(device_index=0)
-    memory_monitor.start()
-    try:
-        with OmniRunner(model, **omni_kwargs) as runner:
-            generator = torch.Generator(
-                device=current_omni_platform.device_type,
-            ).manual_seed(seed)
-            outputs = runner.omni.generate(
-                "a photo of a cat sitting on a laptop keyboard",
-                OmniDiffusionSamplingParams(
-                    height=height,
-                    width=width,
-                    num_inference_steps=num_inference_steps,
-                    guidance_scale=0.0,
-                    generator=generator,
-                ),
-            )
-    finally:
-        memory_monitor.stop()
-    peak_mem = memory_monitor.peak_used_mb / 1024
-
-    first_output = outputs[0]
-    assert first_output.final_output_type == "image"
-    if hasattr(first_output, "images") and first_output.images:
-        images = first_output.images
-    else:
-        assert hasattr(first_output, "request_output") and first_output.request_output
-        request_output = first_output.request_output
-        if isinstance(request_output, list):
-            req_out = request_output[0]
-        else:
-            req_out = request_output
-        assert isinstance(req_out, OmniRequestOutput) and hasattr(req_out, "images")
-        images = req_out.images
-    assert len(images) >= 1
-    assert images[0].width == width
-    assert images[0].height == height
-
-    return images, peak_mem
-
-
-@hardware_test(res={"cuda": "L4"})
-@pytest.mark.full_model
-@pytest.mark.diffusion
-@pytest.mark.slow
-@pytest.mark.parametrize("model", [Z_IMAGE_CONFIG, FLUX_CONFIG], ids=["Z-Image-Turbo", "FLUX.2-klein"])
-def test_single_stage_diffusion_gguf(model: DiffusionGGUFTestConfig, monkeypatch: pytest.MonkeyPatch) -> None:
-    """Z-Image-Turbo GGUF generates valid images and uses less memory than BF16."""
-    monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
-
-    # BF16 baseline
-    hf_images, mem_bf16 = _generate_single_stage_image(
-        model=model.hf_model,
-    )
-
-    # GGUF
-    images, mem_gguf = _generate_single_stage_image(
-        model=model.hf_model,
-        diffusion_quantization_config={
-            "method": "gguf",
-            "gguf_model": model.gguf_model,
-        },
-    )
-
-    hf_image_path = f"test_{model.artifact_prefix}_hf.png"
-    gguf_image_path = f"test_{model.artifact_prefix}_gguf.png"
-    assert len(hf_images) >= 1
-    hf_images[0].save(hf_image_path)
-    assert len(images) >= 1
-    images[0].save(gguf_image_path)
-    print(f"Saved HF image: {hf_image_path}")
-    print(f"Saved GGUF image: {gguf_image_path}")
-
-    image_cosine_similarity = _image_cosine_similarity(hf_images[0], images[0])
-    print(f"{model.artifact_prefix} image cosine similarity: {image_cosine_similarity:.4f}")
-    assert image_cosine_similarity >= model.min_cosine_similarity, (
-        f"GGUF image cosine similarity ({image_cosine_similarity:.4f}) should be >= {model.min_cosine_similarity:.4f}"
-    )
-
-    print(f"{model.artifact_prefix} BF16 peak VRAM delta: {mem_bf16:.2f} GiB")
-    print(f"{model.artifact_prefix} GGUF peak VRAM delta: {mem_gguf:.2f} GiB")
-    reduction = (mem_bf16 - mem_gguf) / mem_bf16 * 100
-    print(f"VRAM reduction: {reduction:.1f}%")
-    assert mem_gguf < mem_bf16, f"GGUF ({mem_gguf:.2f} GiB) should use less VRAM than BF16 ({mem_bf16:.2f} GiB)"

tests/test_config_factory.py

@@ -1166,10 +1166,7 @@ class TestQwen3OmniPipeline:
         assert isinstance(s, StagePipelineConfig)
         assert s.input_sources == (0,)
         assert s.sampling_constraints["stop_token_ids"] == [2150]
-        # thinker2talker was removed: sync_process_input_func always wins.
+        assert s.custom_process_input_func is not None
-        assert s.custom_process_input_func is None
-        assert s.sync_process_input_func is not None
-        assert s.sync_process_input_func.endswith("thinker2talker_token_only")
         assert s.custom_process_next_stage_input_func is not None
 
     def test_code2wav(self):
@@ -1183,7 +1180,7 @@ class TestQwen3OmniPipeline:
         assert isinstance(s, StagePipelineConfig)
         assert s.execution_type == StageExecutionType.LLM_GENERATION
         assert s.final_output_type == "audio"
-        assert s.engine_output_type == "audio"
+        assert s.custom_process_input_func is not None
 
 
 class TestQwen2_5OmniPipeline:
@@ -1721,7 +1718,7 @@ class TestAuraOmniDeploy:
     def test_aura_omni_deploy_resolves_four_native_stages(self):
         pipeline_cfg = StageConfigFactory.resolve_pipeline_config("aura_omni")
 
-        stages, _ = StageConfigFactory._create_from_registry(
+        stages = StageConfigFactory._create_from_registry(
             "qwen3_tts",
             pipeline_cfg,
             cli_overrides={},
@@ -1943,12 +1940,11 @@ class TestSentinelDefaultPrecedence:
             model_type,
             Q3_OMNI_ALL_STAGES_HF_CONFIG,
         )
-        stages, _ = StageConfigFactory._create_from_registry(
+        return StageConfigFactory._create_from_registry(
             "qwen3_omni_moe",
             pipeline_cfg,
             cli_overrides=cli_overrides,
         )
-        return stages
 
     def test_typed_kwarg_overrides_yaml(self):
         stages = self._stages({"max_num_seqs": 999})

tests/test_version.py

@@ -56,26 +56,3 @@ def test_warning_contains_version_strings():
     msg = str(record[0].message)
     assert "0.19.0rc2.dev21" in msg
     assert "0.18.0" in msg
-
-
-def test_omni_plugin_loader_invokes_vllm_general_plugins_first(monkeypatch):
-    import vllm.plugins as vllm_plugins
-
-    import vllm_omni.plugins as omni_plugins
-
-    calls = []
-    monkeypatch.setattr(omni_plugins, "omni_plugins_loaded", False)
-    monkeypatch.setattr(vllm_plugins, "load_general_plugins", lambda: calls.append("vllm"))
-    monkeypatch.setattr(
-        omni_plugins,
-        "load_omni_plugins_by_group",
-        lambda group: {"omni": lambda: calls.append(group)},
-    )
-
-    omni_plugins.load_omni_general_plugins()
-
-    assert calls == ["vllm", omni_plugins.OMNI_DEFAULT_PLUGINS_GROUP]
-
-    omni_plugins.load_omni_general_plugins()
-
-    assert calls == ["vllm", omni_plugins.OMNI_DEFAULT_PLUGINS_GROUP]

tests/utils/test_tracking_parser.py

@@ -591,7 +591,7 @@ def test_explicit_cli_arg_reaches_runtime_overrides(mock_stages):
     ns = p.parse_args(["--max-num-seqs", "999"])
 
     explicit_kwargs = ns.get_explicit_kwargs_dict()
-    stages, _ = StageConfigFactory._create_from_registry(
+    stages = StageConfigFactory._create_from_registry(
         _TEST_MODEL,
         _TEST_PIPELINE,
         explicit_kwargs,
@@ -608,7 +608,7 @@ def test_omitted_default_not_in_runtime_overrides(mock_stages):
     ns = p.parse_args([])
 
     explicit_kwargs = ns.get_explicit_kwargs_dict()
-    stages, _ = StageConfigFactory._create_from_registry(
+    stages = StageConfigFactory._create_from_registry(
         _TEST_MODEL,
         _TEST_PIPELINE,
         explicit_kwargs,
@@ -630,7 +630,7 @@ def test_config_file_args_reach_runtime_overrides(mock_stages):
         ns = p.parse_args(["--config", "fake.yaml"])
 
     explicit_kwargs = ns.get_explicit_kwargs_dict()
-    stages, _ = StageConfigFactory._create_from_registry(
+    stages = StageConfigFactory._create_from_registry(
         _TEST_MODEL,
         _TEST_PIPELINE,
         explicit_kwargs,
@@ -648,7 +648,7 @@ def test_per_stage_override_routes_correctly(mock_stages):
     ns = p.parse_args(["--stage-0-gpu-memory-utilization", "0.42"])
 
     explicit_kwargs = ns.get_explicit_kwargs_dict()
-    stages, _ = StageConfigFactory._create_from_registry(
+    stages = StageConfigFactory._create_from_registry(
         _TEST_MODEL,
         _TEST_PIPELINE,
         explicit_kwargs,
@@ -670,7 +670,7 @@ def test_explicit_args_omitted_from_yaml(mock_stages):
     ns = p.parse_args(["--enforce-eager"])
 
     explicit_kwargs = ns.get_explicit_kwargs_dict()
-    stages, _ = StageConfigFactory._create_from_registry(
+    stages = StageConfigFactory._create_from_registry(
         _TEST_MODEL,
         _TEST_PIPELINE,
         explicit_kwargs,

vllm_omni/config/composable_parallel/__init__.py

@@ -1,104 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Composable parallel strategies for vLLM-Omni.
-
-A small, runtime-agnostic vocabulary for declaring a stage's parallel layout
-(:class:`StrategySpec` + mesh/routing/aggregation patterns), plus the logic to
-translate a per-role stack of specs into concrete engine sizing and overlay it
-onto merged stage configs.
-
-The spec/pattern *types* are deliberately pure data with no engine imports; the
-translation and apply logic lives alongside them and consumes vLLM-Omni's own
-config objects.
-"""
-
-from vllm_omni.config.composable_parallel.aggregation import (
-    AGGREGATION_PATTERN_VARIANTS,
-    AggregationConflictError,
-    AggregationPattern,
-    AllGather,
-    Combine,
-    FanInByStage,
-    GatherDim,
-    StitchPipeline,
-    StitchSpatial,
-    TakeRank,
-    Union,
-)
-from vllm_omni.config.composable_parallel.apply import (
-    StrategyApplyError,
-    StrategyApplyResult,
-    apply_strategy_specs,
-    check_device_layout,
-)
-from vllm_omni.config.composable_parallel.routing import (
-    ROUTING_PATTERN_VARIANTS,
-    Broadcast,
-    DuplicateWithCondTag,
-    PartitionByHash,
-    PipelineMicrobatch,
-    RouteByExpert,
-    RouteByStage,
-    RoutingKeyError,
-    RoutingPattern,
-    ShardSequence,
-    ShardSpatial,
-)
-from vllm_omni.config.composable_parallel.spec import (
-    MESH_AXIS_KINDS,
-    KernelSpec,
-    LayerHookSpec,
-    MeshAxisKind,
-    MeshAxisSpec,
-    SpecMergeConflictError,
-    StrategySpec,
-)
-from vllm_omni.config.composable_parallel.translator import (
-    AxisTranslationError,
-    OmniParallelConfig,
-    translate_strategy_stack,
-)
-
-__all__ = [
-    # spec types
-    "MeshAxisKind",
-    "MESH_AXIS_KINDS",
-    "MeshAxisSpec",
-    "LayerHookSpec",
-    "KernelSpec",
-    "StrategySpec",
-    "SpecMergeConflictError",
-    # routing patterns
-    "RoutingPattern",
-    "RoutingKeyError",
-    "Broadcast",
-    "PartitionByHash",
-    "ShardSequence",
-    "ShardSpatial",
-    "RouteByExpert",
-    "PipelineMicrobatch",
-    "DuplicateWithCondTag",
-    "RouteByStage",
-    "ROUTING_PATTERN_VARIANTS",
-    # aggregation patterns
-    "AggregationPattern",
-    "AggregationConflictError",
-    "TakeRank",
-    "Union",
-    "GatherDim",
-    "AllGather",
-    "StitchSpatial",
-    "StitchPipeline",
-    "Combine",
-    "FanInByStage",
-    "AGGREGATION_PATTERN_VARIANTS",
-    # translation
-    "OmniParallelConfig",
-    "translate_strategy_stack",
-    "AxisTranslationError",
-    # apply
-    "apply_strategy_specs",
-    "StrategyApplyResult",
-    "check_device_layout",
-    "StrategyApplyError",
-]

vllm_omni/config/composable_parallel/aggregation.py

@@ -1,126 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Closed aggregation-pattern hierarchy for declarative parallel strategies.
-
-An aggregation pattern describes how the per-worker results of one mesh axis
-are recombined into a single logical result (e.g. take one tensor-parallel
-rank's output, union disjoint data-parallel shards). The hierarchy is *closed*:
-subclasses may only be defined in this module.
-"""
-
-from __future__ import annotations
-
-from collections.abc import Callable
-from dataclasses import dataclass
-
-
-@dataclass(frozen=True)
-class AggregationPattern:
-    """Base class for closed aggregation patterns. Subclass only within this module."""
-
-    def __init_subclass__(cls, **kwargs: object) -> None:
-        super().__init_subclass__(**kwargs)
-        if cls.__module__ != __name__:
-            raise TypeError("AggregationPattern is closed to external subclassing")
-
-    def pattern_kind(self) -> str:
-        raise NotImplementedError
-
-
-class AggregationConflictError(ValueError):
-    """Raised when two aggregation inputs conflict for the same request id."""
-
-
-@dataclass(frozen=True)
-class TakeRank(AggregationPattern):
-    """Select one rank's result (TP — all ranks agree)."""
-
-    rank: int = 0
-
-    def __post_init__(self) -> None:
-        if not isinstance(self.rank, int) or isinstance(self.rank, bool):
-            raise TypeError(f"TakeRank.rank must be int, got {type(self.rank).__name__}")
-
-    def pattern_kind(self) -> str:
-        return "take_rank"
-
-
-@dataclass(frozen=True)
-class Union(AggregationPattern):
-    """Union of disjoint replica results (DP, HSDP)."""
-
-    def pattern_kind(self) -> str:
-        return "union"
-
-
-@dataclass(frozen=True)
-class GatherDim(AggregationPattern):
-    """Gather along a dimension and materialize on one designated rank.
-
-    Default designated rank is axis-rank 0 unless adapter overrides.
-    """
-
-    dim: int = 1
-
-    def pattern_kind(self) -> str:
-        return "gather_dim"
-
-
-@dataclass(frozen=True)
-class AllGather(AggregationPattern):
-    """All ranks materialize the same aggregate result across the axis."""
-
-    dim: int | None = None
-
-    def pattern_kind(self) -> str:
-        return "all_gather"
-
-
-@dataclass(frozen=True)
-class StitchSpatial(AggregationPattern):
-    """Stitch spatial patches (VAE-PP)."""
-
-    grid: tuple[int, int] = (1, 1)
-
-    def pattern_kind(self) -> str:
-        return "stitch_spatial"
-
-
-@dataclass(frozen=True)
-class StitchPipeline(AggregationPattern):
-    """Stitch pipeline stage outputs (PP)."""
-
-    def pattern_kind(self) -> str:
-        return "stitch_pipeline"
-
-
-@dataclass(frozen=True)
-class Combine(AggregationPattern):
-    """Custom reducer combination (CFG). Adapter resolves reducer_id."""
-
-    reducer_id: str = "default"
-
-    def pattern_kind(self) -> str:
-        return "combine"
-
-
-@dataclass(frozen=True)
-class FanInByStage(AggregationPattern):
-    """Fan-in results routed by stage (stage replicas)."""
-
-    def pattern_kind(self) -> str:
-        return "fan_in_by_stage"
-
-
-AGGREGATION_PATTERN_VARIANTS: tuple[type[AggregationPattern], ...] = (
-    TakeRank,
-    Union,
-    GatherDim,
-    AllGather,
-    StitchSpatial,
-    StitchPipeline,
-    Combine,
-    FanInByStage,
-)
-
-ReducerFn = Callable[..., object]

vllm_omni/config/composable_parallel/apply.py

@@ -1,264 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Apply declarative parallel strategies onto merged stage configs.
-
-This is the *override-after-merge* seam. ``merge_pipeline_deploy`` first fuses
-the pipeline topology + deploy YAML into a list of ``StageConfig`` objects; this
-module then overlays the engine sizing derived from a per-role
-``StrategySpec`` stack onto those stages.
-
-Design rules (one writer per axis):
-
-* A strategy only writes the axes it actually declares. If a stack has no ``tp``
-  axis, this module never touches ``tensor_parallel_size`` — it does not force a
-  default of 1.
-* **Conflict-on-explicit.** If a knob was already set explicitly by the deploy
-  YAML and the strategy derives a different value, we raise rather than silently
-  override. Equal values are a no-op; unset values are filled.
-* **Pre-spawn device check.** When a stage declares ``devices``, the device
-  count must be consistent with the engine world size (``tp * dp * pp``) and the
-  replica count, so misconfigurations fail at config time instead of at spawn.
-
-``omni_lb_policy`` is *not* a per-stage config knob: omni reads it once at
-orchestrator construction (``AsyncOmniEngine``), not from stage configs. So a
-stage_replica axis's derived policy is surfaced on :class:`StrategyApplyResult`
-for the caller to pass to the orchestrator; it is never silently written into a
-stage where it would be ignored.
-"""
-
-from __future__ import annotations
-
-from collections.abc import Mapping, Sequence
-from dataclasses import dataclass, field
-from typing import Any, NoReturn
-
-from vllm.logger import init_logger
-
-from vllm_omni.config.composable_parallel.spec import StrategySpec
-from vllm_omni.config.composable_parallel.translator import (
-    OmniParallelConfig,
-    translate_strategy_stack,
-)
-
-logger = init_logger(__name__)
-
-# Role key -> the StageConfig fields each declared axis writes.
-_ENGINE_FIELD_BY_KIND: dict[str, str] = {
-    "tp": "tensor_parallel_size",
-    "dp": "data_parallel_size",
-    "pp": "pipeline_parallel_size",
-}
-
-RoleKey = str
-
-
-class StrategyApplyError(ValueError):
-    """Error for applying a strategy onto stage configs.
-
-    A single error type (rather than per-cause subclasses) keeps the public
-    surface small and consistent with the rest of the codebase; the specific
-    cause — an unmatched role, a conflict with an explicit deploy value, or a
-    device-count mismatch — is in the message and is logged before the raise so
-    it stays visible even when the type is unavailable to a caller (e.g. across a
-    server boundary).
-    """
-
-
-def _fail(msg: str) -> NoReturn:
-    """Log and raise :class:`StrategyApplyError`."""
-    logger.error("[composable_parallel] %s", msg)
-    raise StrategyApplyError(msg)
-
-
-@dataclass
-class StrategyApplyResult:
-    """Outcome of applying strategies to a stage list.
-
-    ``stages`` is the same list passed in (mutated in place and returned for
-    convenience). ``omni_lb_policy`` is the pipeline-wide load-balancer policy
-    derived from any stage_replica axes (``None`` if none declared one); the
-    caller wires it into the orchestrator, since omni does not read it from
-    stage configs.
-
-    ``per_role_config`` is keyed by the role key the caller supplied — a
-    ``model_stage`` name (str). ``per_stage_config`` is the same information
-    keyed by the resolved integer ``stage_id`` — handy for re-validating a
-    stage after later layers (e.g. CLI overrides) have run.
-    """
-
-    stages: list[Any]
-    omni_lb_policy: str | None = None
-    per_role_config: dict[RoleKey, OmniParallelConfig] = field(default_factory=dict)
-    per_stage_config: dict[int, OmniParallelConfig] = field(default_factory=dict)
-
-
-def _resolve_stage(stages: Sequence[Any], key: RoleKey) -> Any:
-    """Find the single stage matching ``key`` by ``model_stage`` name.
-
-    Role keys are ``model_stage`` names (str) — the same human-readable names
-    that strategy.yaml authors use to refer to a stage (e.g. ``"thinker"``).
-    """
-    matches = [s for s in stages if getattr(s, "model_stage", None) == key]
-    descriptor = f"model_stage={key!r}"
-
-    if not matches:
-        available = ", ".join(f"{getattr(s, 'model_stage', '?')!r}(id={getattr(s, 'stage_id', '?')})" for s in stages)
-        _fail(f"strategy role {descriptor} did not match any stage; available stages: {available}")
-    if len(matches) > 1:
-        _fail(f"strategy role {descriptor} is ambiguous; it matched {len(matches)} stages")
-    return matches[0]
-
-
-def _set_explicit(engine_args: dict[str, Any], field_name: str, value: Any, *, role: RoleKey) -> None:
-    """Write ``field_name`` with conflict-on-explicit semantics.
-
-    ``engine_args`` is the raw mapping loaded straight from the deploy YAML
-    (before the engine-args dataclass is constructed), so a key is *present* only
-    when the YAML set it explicitly. We therefore key off membership rather than
-    ``.get``: an explicit ``None`` in the YAML is a real, present value and must
-    not be conflated with a key miss (which ``.get`` would report as ``None``).
-    """
-    if field_name in engine_args and engine_args[field_name] != value:
-        existing = engine_args[field_name]
-        _fail(
-            f"role {role!r}: strategy derives {field_name}={value} but the deploy config already "
-            f"set {field_name}={existing}. Remove one of them (the strategy is the single writer "
-            "for the axes it declares)."
-        )
-    engine_args[field_name] = value
-
-
-def _set_num_replicas(runtime: dict[str, Any], value: int, *, role: RoleKey) -> None:
-    """Write ``num_replicas`` with conflict-on-explicit (treating 1 as unset)."""
-    existing = runtime.get("num_replicas", 1)
-    if existing not in (1, value):
-        _fail(
-            f"role {role!r}: strategy derives num_replicas={value} but the deploy config already "
-            f"set num_replicas={existing}. Remove one of them."
-        )
-    runtime["num_replicas"] = value
-
-
-def _parse_device_count(devices: Any) -> int | None:
-    """Return the number of device ids in a ``devices`` value, or None."""
-    if devices is None:
-        return None
-    if isinstance(devices, (list, tuple)):
-        return len([d for d in devices])
-    text = str(devices).strip()
-    if not text:
-        return None
-    return len([d for d in text.split(",") if d.strip()])
-
-
-def check_device_layout(
-    devices: Any,
-    *,
-    tensor_parallel_size: int,
-    data_parallel_size: int,
-    pipeline_parallel_size: int,
-    num_replicas: int,
-    role: RoleKey,
-) -> None:
-    """Validate a stage's device count against its world size (× replicas in pool mode).
-
-    Raises :class:`StrategyApplyError` when the declared ``devices`` count is
-    neither the per-replica world size nor the full ``num_replicas`` pool.
-    Exposed publicly so the resolved (post-CLI) layout can be re-checked after
-    later override layers, not just the strategy-derived snapshot.
-    """
-    count = _parse_device_count(devices)
-    if count is None:
-        return
-    world = int(tensor_parallel_size) * int(data_parallel_size) * int(pipeline_parallel_size)
-    replicas = int(num_replicas or 1)
-    # Two valid shapes: a single per-replica template (== world) or the full
-    # pool across replicas (== replicas * world). This mirrors omni's own
-    # per-replica device splitter.
-    valid = {world, replicas * world}
-    if count not in valid:
-        _fail(
-            f"role {role!r}: declared {count} device id(s) but the strategy world size is {world} "
-            f"(tp={tensor_parallel_size} * dp={data_parallel_size} * pp={pipeline_parallel_size}). "
-            f"Provide either {world} (per-replica template) or {replicas * world} "
-            f"(num_replicas={replicas} pool)."
-        )
-
-
-def _check_devices(runtime: dict[str, Any], cfg: OmniParallelConfig, *, role: RoleKey) -> None:
-    """Pre-spawn check on the strategy-derived snapshot (see :func:`check_device_layout`)."""
-    check_device_layout(
-        runtime.get("devices"),
-        tensor_parallel_size=cfg.tensor_parallel_size,
-        data_parallel_size=cfg.data_parallel_size,
-        pipeline_parallel_size=cfg.pipeline_parallel_size,
-        num_replicas=int(runtime.get("num_replicas", 1) or 1),
-        role=role,
-    )
-
-
-def _apply_to_stage(stage: Any, cfg: OmniParallelConfig, *, role: RoleKey) -> None:
-    engine_args = stage.yaml_engine_args
-    runtime = stage.yaml_runtime
-
-    declared = set(cfg.l1_owners.keys())
-    for kind, field_name in _ENGINE_FIELD_BY_KIND.items():
-        if kind in declared:
-            _set_explicit(engine_args, field_name, getattr(cfg, field_name), role=role)
-    if "ep" in declared and cfg.enable_expert_parallel:
-        _set_explicit(engine_args, "enable_expert_parallel", True, role=role)
-    if "stage_replica" in declared:
-        _set_num_replicas(runtime, cfg.stage_replica_size, role=role)
-
-    _check_devices(runtime, cfg, role=role)
-
-
-def apply_strategy_specs(
-    stages: list[Any],
-    strategy_specs: Mapping[RoleKey, Sequence[StrategySpec]],
-) -> StrategyApplyResult:
-    """Overlay per-role strategy specs onto a merged stage list.
-
-    Args:
-        stages: the ``list[StageConfig]`` returned by ``merge_pipeline_deploy``.
-        strategy_specs: maps a role (a ``model_stage`` name, e.g. ``"thinker"``)
-            to that stage's stack of ``StrategySpec`` (one per declared mesh
-            axis). Role keys are ``model_stage`` names (str).
-
-    Returns:
-        A :class:`StrategyApplyResult` with the mutated stages and the derived
-        pipeline-wide ``omni_lb_policy`` (if any stage_replica axis set one).
-
-    Raises:
-        StrategyApplyError: a role matched zero or multiple stages, a derived
-            value conflicts with an explicit deploy value (including two roles
-            deriving different ``omni_lb_policy`` values), or a stage's device
-            count is inconsistent with its world size.
-        AxisTranslationError: the spec stack is invalid/unsupported.
-        NotImplementedError: the spec stack requests routing not built yet.
-    """
-    result = StrategyApplyResult(stages=stages)
-    lb_policy: str | None = None
-    lb_owner: RoleKey | None = None
-
-    for key, specs in strategy_specs.items():
-        stage = _resolve_stage(stages, key)
-        cfg = translate_strategy_stack(specs)
-        _apply_to_stage(stage, cfg, role=key)
-        result.per_role_config[key] = cfg
-        stage_id = getattr(stage, "stage_id", None)
-        if stage_id is not None:
-            result.per_stage_config[int(stage_id)] = cfg
-
-        if cfg.stage_replica_size > 1 and cfg.omni_lb_policy is not None:
-            if lb_policy is not None and lb_policy != cfg.omni_lb_policy:
-                _fail(
-                    f"role {key!r} derives omni_lb_policy={cfg.omni_lb_policy!r} but role "
-                    f"{lb_owner!r} already derived {lb_policy!r}. The load-balancer policy is "
-                    "pipeline-wide; only one value is allowed."
-                )
-            lb_policy = cfg.omni_lb_policy
-            lb_owner = key
-
-    result.omni_lb_policy = lb_policy
-    return result

vllm_omni/config/composable_parallel/routing.py

@@ -1,138 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Closed routing-pattern hierarchy for declarative parallel strategies.
-
-A routing pattern describes how a request batch is distributed across the
-workers of one mesh axis (e.g. broadcast to every tensor-parallel rank, hash
-partition across data-parallel ranks). The hierarchy is *closed*: subclasses
-may only be defined in this module, so the set of patterns is a fixed,
-exhaustively matchable contract.
-"""
-
-from __future__ import annotations
-
-from collections.abc import Callable
-from dataclasses import dataclass
-from typing import Literal
-
-
-@dataclass(frozen=True)
-class RoutingPattern:
-    """Base class for closed routing patterns. Subclass only within this module."""
-
-    def __init_subclass__(cls, **kwargs: object) -> None:
-        super().__init_subclass__(**kwargs)
-        if cls.__module__ != __name__:
-            raise TypeError("RoutingPattern is closed to external subclassing")
-
-    def pattern_kind(self) -> str:
-        raise NotImplementedError
-
-
-class RoutingKeyError(KeyError):
-    """Raised when a routing key is missing from a request."""
-
-    def __init__(self, request_id: object, missing_key: str) -> None:
-        self.request_id = request_id
-        self.missing_key = missing_key
-        super().__init__(f"Missing routing key {missing_key!r} for request_id={request_id!r}")
-
-
-@dataclass(frozen=True)
-class Broadcast(RoutingPattern):
-    """Every worker in the group sees the whole batch (TP, dense EP)."""
-
-    def pattern_kind(self) -> str:
-        return "broadcast"
-
-
-@dataclass(frozen=True)
-class PartitionByHash(RoutingPattern):
-    """Stable hash partition by request key (DP, HSDP)."""
-
-    key: str = "request_id"
-
-    def pattern_kind(self) -> str:
-        return "partition_by_hash"
-
-
-@dataclass(frozen=True)
-class ShardSequence(RoutingPattern):
-    """Shard along a sequence dimension (SP-Ulysses, SP-Ring, CP)."""
-
-    dim: int = 1
-
-    def pattern_kind(self) -> str:
-        return "shard_sequence"
-
-
-@dataclass(frozen=True)
-class ShardSpatial(RoutingPattern):
-    """Shard along a spatial grid (VAE-PP)."""
-
-    grid: tuple[int, int] = (1, 1)
-
-    def pattern_kind(self) -> str:
-        return "shard_spatial"
-
-
-@dataclass(frozen=True)
-class RouteByExpert(RoutingPattern):
-    """Sparse MoE expert routing."""
-
-    router_id: str = "default"
-
-    def pattern_kind(self) -> str:
-        return "route_by_expert"
-
-
-@dataclass(frozen=True)
-class PipelineMicrobatch(RoutingPattern):
-    """Pipeline parallelism microbatch routing (PP)."""
-
-    microbatch_size: int = 1
-
-    def pattern_kind(self) -> str:
-        return "pipeline_microbatch"
-
-
-@dataclass(frozen=True)
-class DuplicateWithCondTag(RoutingPattern):
-    """Duplicate batch with a conditional tag (CFG)."""
-
-    flag_name: str = "guidance_branch"
-
-    def pattern_kind(self) -> str:
-        return "duplicate_with_cond_tag"
-
-
-@dataclass(frozen=True)
-class RouteByStage(RoutingPattern):
-    """Route requests by pipeline stage policy (stage replicas).
-
-    The ``random``/``round_robin``/``least_queue`` policies describe stateless
-    load balancing that a runtime (e.g. an engine's built-in DP load balancer,
-    or the omni StagePool balancer) can realize. ``hash`` describes
-    deterministic, key-stable routing that a load balancer cannot honor on its
-    own (see ``PartitionByHash``).
-    """
-
-    routing_policy: Literal["random", "round_robin", "least_queue", "hash"] = "round_robin"
-
-    def pattern_kind(self) -> str:
-        return "route_by_stage"
-
-
-ROUTING_PATTERN_VARIANTS: tuple[type[RoutingPattern], ...] = (
-    Broadcast,
-    PartitionByHash,
-    ShardSequence,
-    ShardSpatial,
-    RouteByExpert,
-    PipelineMicrobatch,
-    DuplicateWithCondTag,
-    RouteByStage,
-)
-
-# Optional callable slot for Combine-style custom routing reducers in adapters.
-RouterFn = Callable[..., object]

vllm_omni/config/composable_parallel/spec.py

@@ -1,136 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Core declarative types for parallel-strategy specification.
-
-A :class:`StrategySpec` declares one parallelism scheme: the mesh axis it sizes
-(kind + degree), how a batch is routed across that axis, and how per-worker
-results are aggregated back. A *stack* of specs (one per axis) fully describes a
-stage's parallel layout in a runtime-agnostic, data-only form that can be
-translated into concrete engine sizing.
-
-The mesh-axis kinds below enumerate every parallelism dimension the contract can
-describe. Only a subset is wired end-to-end today (see ``translator.py``):
-
-* **Wired** (translatable now): ``tp``, ``dp``, ``pp``, ``ep``,
-  ``stage_replica``.
-* **Reserved** (declared in the type system but not yet translatable — the
-  translator raises ``AxisTranslationError`` for them): ``sp_ulysses``,
-  ``sp_ring``, ``cfg``, ``vae_pp``, ``hsdp``, ``stage_pp``, ``cp``.
-
-Reserved kinds fail fast at translation time rather than silently doing nothing,
-so declaring one is an explicit "not yet" rather than a no-op.
-"""
-
-from __future__ import annotations
-
-from collections.abc import Mapping
-from dataclasses import dataclass, field
-from enum import Enum
-from typing import Any, Literal
-
-from vllm_omni.config.composable_parallel.aggregation import AggregationPattern
-from vllm_omni.config.composable_parallel.routing import RoutingPattern
-
-
-class MeshAxisKind(str, Enum):
-    """Enumerates every parallelism dimension the strategy contract can describe.
-
-    Subclassing ``str`` keeps existing string comparisons and dict keys working
-    (``MeshAxisKind.TP == "tp"`` and both hash the same), while making the type
-    refactor-safe: a typo'd kind is an ``AttributeError`` on the enum rather than
-    a silently-wrong bare string.
-    """
-
-    TP = "tp"
-    DP = "dp"
-    PP = "pp"
-    EP = "ep"
-    SP_ULYSSES = "sp_ulysses"
-    SP_RING = "sp_ring"
-    CFG = "cfg"
-    VAE_PP = "vae_pp"
-    HSDP = "hsdp"
-    STAGE_PP = "stage_pp"
-    STAGE_REPLICA = "stage_replica"
-    CP = "cp"
-
-
-# Bare string values, kept as a plain tuple so membership checks and error
-# messages read as the kind names (``"tp"``) rather than enum reprs. ``str``-enum
-# equality means both ``"tp"`` and ``MeshAxisKind.TP`` satisfy ``in`` checks.
-MESH_AXIS_KINDS: tuple[str, ...] = tuple(k.value for k in MeshAxisKind)
-
-HookCategory = Literal[
-    "linear",
-    "attention_pre",
-    "attention_post",
-    "ffn_pre",
-    "ffn_post",
-    "moe_dispatch",
-    "other",
-]
-
-HOOK_CATEGORY_ORDER: dict[HookCategory, int] = {
-    "linear": 0,
-    "attention_pre": 1,
-    "attention_post": 2,
-    "ffn_pre": 3,
-    "ffn_post": 4,
-    "moe_dispatch": 5,
-    "other": 6,
-}
-
-
-class SpecMergeConflictError(ValueError):
-    """Raised when merged hook/kernel specs have conflicting declarations."""
-
-
-@dataclass(frozen=True)
-class MeshAxisSpec:
-    """Declares one axis of a process mesh (kind + size)."""
-
-    kind: MeshAxisKind
-    size: int
-
-    def __post_init__(self) -> None:
-        if self.kind not in MESH_AXIS_KINDS:
-            raise ValueError(f"MeshAxisSpec.kind must be one of {MESH_AXIS_KINDS}, got {self.kind!r}")
-        if self.size <= 0:
-            raise ValueError(f"MeshAxisSpec.size must be > 0, got {self.size}")
-
-
-@dataclass(frozen=True)
-class LayerHookSpec:
-    """L2 hook slot referenced by StrategySpec; resolved by the model walker."""
-
-    hook_id: str
-    target: str | None = None
-    category: HookCategory = "other"
-    priority: int = 0
-    axis_index: int = 0
-
-
-@dataclass(frozen=True)
-class KernelSpec:
-    """L3 kernel slot referenced by StrategySpec; resolved by the kernel registry."""
-
-    kernel_id: str
-    target: str | None = None
-    category: HookCategory = "other"
-    priority: int = 0
-    axis_index: int = 0
-    group_axis_kind: MeshAxisKind | None = None
-    requires_collective: bool = False
-
-
-@dataclass(frozen=True)
-class StrategySpec:
-    """Declarative contract for one parallelism scheme (data + hook/kernel slots)."""
-
-    name: str
-    mesh_axis: MeshAxisSpec
-    routing: RoutingPattern
-    aggregation: AggregationPattern
-    layer_hook_specs: tuple[LayerHookSpec, ...] = ()
-    kernel_specs: tuple[KernelSpec, ...] = ()
-    shard_extension: Mapping[str, Any] = field(default_factory=dict)

vllm_omni/config/composable_parallel/strategy_loader.py

@@ -1,149 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Load per-role parallel strategies from a strategy file.
-
-The strategy file is a small, optional companion to the deploy config. It maps a
-logical role (``model_stage``, e.g. ``thinker``) to a list of axis declarations,
-each of which becomes one :class:`StrategySpec`. YAML is the only format wired up
-today, but the module/symbol names stay format-agnostic so a different source can
-be added without renaming the public surface::
-
-    strategies:
-      thinker:
-        - axis: tp
-          size: 2
-      talker:
-        - axis: stage_replica
-          size: 2
-          routing: round_robin   # random | round_robin | least_queue
-      code2wav:
-        - axis: stage_replica
-          size: 2
-          routing: round_robin
-
-Per-axis fields:
-
-* ``axis`` (required): mesh-axis kind (``tp``, ``dp``, ``pp``, ``ep``,
-  ``stage_replica`` are translatable today).
-* ``size`` (required): axis degree (> 0).
-* ``routing`` (optional): stateless policy for ``dp`` / ``stage_replica`` axes
-  (``random`` / ``round_robin`` / ``least_queue``). Ignored for axes whose
-  routing is fixed (``tp`` broadcast, ``pp`` microbatch).
-* ``l1_owner`` (optional): ``engine`` or ``delegated``; overrides the default
-  owner for the axis (recorded in ``shard_extension``).
-* ``name`` (optional): a label for the spec; defaults to the axis kind.
-
-Routing/aggregation patterns are filled in from sensible per-kind defaults so
-the file stays compact; the translator validates the combination.
-"""
-
-from __future__ import annotations
-
-from collections.abc import Mapping
-from typing import Any
-
-from vllm_omni.config.composable_parallel.aggregation import (
-    AggregationPattern,
-    FanInByStage,
-    StitchPipeline,
-    TakeRank,
-    Union,
-)
-from vllm_omni.config.composable_parallel.routing import (
-    Broadcast,
-    PipelineMicrobatch,
-    RouteByStage,
-    RoutingPattern,
-)
-from vllm_omni.config.composable_parallel.spec import MeshAxisSpec, StrategySpec
-from vllm_omni.config.yaml_util import load_yaml_config, to_dict
-
-# Per-kind defaults for routing + aggregation, so the file only needs axis+size.
-_ROUTING_POLICY_KINDS = frozenset({"dp", "stage_replica"})
-
-
-class StrategyLoadError(ValueError):
-    """Raised when a strategy file is malformed."""
-
-
-def _default_routing(kind: str, routing_policy: str | None) -> RoutingPattern:
-    if kind in _ROUTING_POLICY_KINDS:
-        return RouteByStage(routing_policy=routing_policy or "round_robin")  # type: ignore[arg-type]
-    if kind == "pp":
-        return PipelineMicrobatch()
-    # tp, ep, and any other kind default to broadcast (the translator rejects
-    # kinds it cannot handle before it ever inspects routing).
-    return Broadcast()
-
-
-def _default_aggregation(kind: str) -> AggregationPattern:
-    if kind == "stage_replica":
-        return FanInByStage()
-    if kind == "pp":
-        return StitchPipeline()
-    if kind in ("dp", "ep"):
-        return Union()
-    return TakeRank()
-
-
-def _build_spec(role: str, entry: Mapping[str, Any]) -> StrategySpec:
-    if "axis" not in entry:
-        raise StrategyLoadError(f"role {role!r}: each strategy entry needs an 'axis' field")
-    if "size" not in entry:
-        raise StrategyLoadError(f"role {role!r}: strategy entry for axis {entry['axis']!r} needs a 'size'")
-
-    kind = str(entry["axis"])
-    try:
-        size = int(entry["size"])
-    except (TypeError, ValueError) as exc:
-        raise StrategyLoadError(f"role {role!r}: axis {kind!r} size must be an integer, got {entry['size']!r}") from exc
-
-    routing_policy = entry.get("routing")
-    if routing_policy is not None and kind not in _ROUTING_POLICY_KINDS:
-        raise StrategyLoadError(
-            f"role {role!r}: axis {kind!r} does not accept a 'routing' policy (only {sorted(_ROUTING_POLICY_KINDS)} do)"
-        )
-
-    shard_extension: dict[str, Any] = {}
-    l1_owner = entry.get("l1_owner")
-    if l1_owner is not None:
-        shard_extension["l1_owner"] = str(l1_owner)
-
-    return StrategySpec(
-        name=str(entry.get("name", kind)),
-        mesh_axis=MeshAxisSpec(kind=kind, size=size),  # type: ignore[arg-type]
-        routing=_default_routing(kind, routing_policy),
-        aggregation=_default_aggregation(kind),
-        shard_extension=shard_extension,
-    )
-
-
-def parse_strategy_specs(data: Mapping[str, Any]) -> dict[str, list[StrategySpec]]:
-    """Parse an already-loaded strategy mapping into per-role spec stacks."""
-    strategies = data.get("strategies", data)
-    if not isinstance(strategies, Mapping):
-        raise StrategyLoadError("strategy file must contain a 'strategies' mapping of role -> list of axes")
-
-    result: dict[str, list[StrategySpec]] = {}
-    for role, entries in strategies.items():
-        if not isinstance(entries, (list, tuple)):
-            raise StrategyLoadError(
-                f"role {role!r}: expected a list of axis declarations, got {type(entries).__name__}"
-            )
-        specs: list[StrategySpec] = []
-        for entry in entries:
-            if not isinstance(entry, Mapping):
-                raise StrategyLoadError(
-                    f"role {role!r}: each axis declaration must be a mapping, got {type(entry).__name__}"
-                )
-            specs.append(_build_spec(str(role), dict(entry)))
-        result[str(role)] = specs
-    return result
-
-
-def load_strategy_specs(path: str) -> dict[str, list[StrategySpec]]:
-    """Load and parse a strategy file into per-role spec stacks."""
-    data = to_dict(load_yaml_config(path))
-    if not isinstance(data, Mapping):
-        raise StrategyLoadError(f"strategy file {path!r} did not parse to a mapping")
-    return parse_strategy_specs(data)

vllm_omni/config/composable_parallel/translator.py

@@ -1,378 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Translate a ``StrategySpec`` stack into vLLM-Omni parallel sizing.
-
-Read a stack of strategy specs (one per mesh axis) and work out the
-tensor/data/pipeline parallel sizes, the dense-EP flag, and the number of stage
-replicas, plus who owns each axis's request routing. The result is a plain,
-CPU-computable sizing struct (:class:`OmniParallelConfig`) keyed by the real
-``OmniEngineArgs`` / ``EngineArgs`` field names, so the deploy layer can splat it
-onto a stage's engine args.
-
-The distinction this module enforces — engine data parallelism (a true vLLM
-intra-engine world dimension) vs. omni stage replicas (independent engines fanned
-out by omni's coordinator) — is documented on the axis validators that depend on
-it (see :func:`_validate_dp` and :func:`_stage_replica_lb_policy`). Routing we do
-not support yet (key-stable / affinity routing) raises ``NotImplementedError``;
-any other invalid spec raises :class:`AxisTranslationError`.
-"""
-
-from __future__ import annotations
-
-from collections.abc import Mapping, Sequence
-from dataclasses import dataclass, field
-from typing import Literal, NoReturn, cast, get_args
-
-from vllm.logger import init_logger
-
-from vllm_omni.config.composable_parallel.routing import (
-    Broadcast,
-    PartitionByHash,
-    PipelineMicrobatch,
-    RouteByStage,
-    RoutingPattern,
-)
-from vllm_omni.config.composable_parallel.spec import MeshAxisKind, StrategySpec
-
-logger = init_logger(__name__)
-
-# Who owns an axis's request routing. A closed string type so an unexpected raw
-# value is caught statically and at runtime (``_VALID_L1_OWNERS`` is derived from
-# it, keeping one source of truth) rather than silently flowing through.
-L1Owner = Literal["delegated", "engine"]
-
-# DP, TP, PP, (dense) EP and stage_replica translate today. DP/TP/PP are true
-# world dimensions vLLM realizes intra-engine; dense EP is a flag over the
-# existing TP*DP ranks; stage_replica is an omni-coordinator-level fan-out of
-# independent engine replicas (NOT a vLLM world dimension). The remaining kinds
-# (sequence/context parallel, CFG, VAE pipelines, ...) need per-layer hooks or
-# custom collectives and arrive in later stages.
-_SUPPORTED_KINDS: tuple[MeshAxisKind, ...] = ("dp", "tp", "pp", "ep", "stage_replica")
-
-# Which EngineArgs field each axis kind *sizes*. EP and stage_replica are
-# deliberately absent: EP is not an independent world dimension but an
-# ``enable_expert_parallel`` flag whose degree equals tp*dp; stage_replica is a
-# per-stage deploy ``num_replicas`` count, not an intra-engine world dimension.
-_AXIS_TO_ENGINE_FIELD: dict[MeshAxisKind, str] = {
-    "tp": "tensor_parallel_size",
-    "dp": "data_parallel_size",
-    "pp": "pipeline_parallel_size",
-}
-
-# Default L1 owner per axis kind. ``stage_replica`` is the only delegated axis:
-# omni's coordinator/StagePool owns routing across replicas. ``dp`` is engine
-# data parallelism, realized by vLLM's own intra-engine DP load balancer.
-_DEFAULT_L1_OWNER: dict[MeshAxisKind, L1Owner] = {
-    "dp": "engine",
-    "tp": "engine",
-    "pp": "engine",
-    "ep": "engine",
-    "stage_replica": "delegated",
-}
-
-# How a RouteByStage policy maps onto omni's load balancer policy string. These
-# are the stateless options omni can actually do; note there's no "hash" here,
-# because omni has no key-stable balancer.
-_STAGE_POLICY_TO_OMNI_LB: dict[str, str] = {
-    "random": "random",
-    "round_robin": "round-robin",
-    "least_queue": "least-queue-length",
-}
-
-_VALID_L1_OWNERS: frozenset[str] = frozenset(get_args(L1Owner))
-
-
-class AxisTranslationError(ValueError):
-    """Error for an invalid or unsupported strategy spec.
-
-    A single error type (rather than a tree of subclasses) keeps the public
-    surface small and consistent with the rest of the codebase; the specific
-    cause is in the message and is logged before the raise so it is visible even
-    when the type is unavailable to a caller (e.g. across a server boundary).
-
-    Strategies that are *valid but not built yet* — key-stable / affinity routing
-    today — raise ``NotImplementedError`` instead, to distinguish "we haven't
-    implemented this" from "your config is wrong".
-    """
-
-
-def _fail(msg: str) -> NoReturn:
-    """Log and raise :class:`AxisTranslationError` for an invalid/unsupported spec."""
-    logger.error("[composable_parallel] %s", msg)
-    raise AxisTranslationError(msg)
-
-
-def _not_implemented(msg: str) -> NoReturn:
-    """Log and raise ``NotImplementedError`` for a valid-but-unbuilt strategy."""
-    logger.error("[composable_parallel] %s", msg)
-    raise NotImplementedError(msg)
-
-
-@dataclass(frozen=True)
-class OmniParallelConfig:
-    """Result of translating a spec stack into omni parallel sizing."""
-
-    tensor_parallel_size: int = 1
-    data_parallel_size: int = 1
-    pipeline_parallel_size: int = 1
-    # Dense expert parallelism: a flag over the TP*DP ranks, not a world axis.
-    enable_expert_parallel: bool = False
-    # Number of independent omni stage replicas (the ``stage_replica`` axis).
-    # This is NOT a vLLM world dimension; it is the per-stage deploy
-    # ``num_replicas`` count. 1 means a single (un-replicated) engine.
-    stage_replica_size: int = 1
-    # The omni StagePool LB policy string for the (delegated) stage_replica
-    # axis, if any. Only meaningful when ``stage_replica_size > 1``.
-    omni_lb_policy: str | None = None
-    # axis kind -> resolved L1 owner ("delegated" | "engine").
-    l1_owners: Mapping[MeshAxisKind, L1Owner] = field(default_factory=dict)
-
-    @property
-    def world_size(self) -> int:
-        # EP and stage_replica are intentionally excluded. EP reuses the TP*DP
-        # ranks rather than adding a dimension; stage_replica spins up separate
-        # engines (each its own world), not extra ranks in this engine's group.
-        # This product is what vLLM calls ``world_size_across_dp`` (tp * pp * dp).
-        return self.tensor_parallel_size * self.data_parallel_size * self.pipeline_parallel_size
-
-    @property
-    def delegated_axes(self) -> tuple[MeshAxisKind, ...]:
-        return tuple(kind for kind, owner in self.l1_owners.items() if owner == "delegated")
-
-    def as_engine_kwargs(self) -> dict[str, object]:
-        """Return per-stage kwargs keyed by real OmniEngineArgs/EngineArgs field names.
-
-        Two derived values are intentionally *not* emitted here because they are
-        not per-stage engine args:
-
-        * ``stage_replica_size`` — a per-stage deploy ``num_replicas`` knob
-          (``StageDeployConfig``); the deploy layer consumes it separately.
-        * ``omni_lb_policy`` — a pipeline-wide load-balancer policy the engine
-          reads once at construction (see ``StrategyApplyResult.omni_lb_policy``);
-          it is applied at the orchestrator level, not folded into per-stage args.
-        """
-        kwargs: dict[str, object] = {
-            "tensor_parallel_size": self.tensor_parallel_size,
-            "data_parallel_size": self.data_parallel_size,
-            "pipeline_parallel_size": self.pipeline_parallel_size,
-        }
-        if self.enable_expert_parallel:
-            kwargs["enable_expert_parallel"] = True
-        return kwargs
-
-
-def _is_affinity_dp_routing(routing: RoutingPattern) -> bool:
-    """True when DP routing demands key-stable (hash) placement."""
-    if isinstance(routing, PartitionByHash):
-        return True
-    if isinstance(routing, RouteByStage) and routing.routing_policy == "hash":
-        return True
-    return False
-
-
-def _resolve_l1_owner(spec: StrategySpec) -> L1Owner:
-    kind = spec.mesh_axis.kind
-    declared_owner = spec.shard_extension.get("l1_owner")
-    owner = declared_owner
-    if owner is None:
-        owner = _DEFAULT_L1_OWNER.get(kind, "engine")
-        if kind not in _DEFAULT_L1_OWNER:
-            logger.debug(
-                "[composable_parallel] axis %r has unknown kind %r with no default l1_owner; falling back to %r",
-                spec.name,
-                kind,
-                owner,
-            )
-        else:
-            logger.debug(
-                "[composable_parallel] axis %r (kind %r) declared no l1_owner; using default %r",
-                spec.name,
-                kind,
-                owner,
-            )
-    if owner not in _VALID_L1_OWNERS:
-        # Distinguish a bad value supplied via shard_extension from a bad default
-        # indexed out of _DEFAULT_L1_OWNER, so the source of the invalid owner is
-        # debuggable rather than swallowed.
-        source = "shard_extension" if declared_owner is not None else "_DEFAULT_L1_OWNER"
-        logger.debug(
-            "[composable_parallel] axis %r (kind %r) resolved invalid l1_owner %r from %s; valid owners are %s",
-            spec.name,
-            kind,
-            owner,
-            source,
-            sorted(_VALID_L1_OWNERS),
-        )
-        _fail(f"axis {kind!r} has invalid l1_owner {owner!r}; expected one of {sorted(_VALID_L1_OWNERS)}")
-    return cast(L1Owner, owner)
-
-
-def _validate_dp(spec: StrategySpec, owner: L1Owner) -> None:
-    """Validate an engine data-parallel axis.
-
-    DP is realized intra-engine by vLLM's own DP load balancer, so it is
-    engine-owned and emits no ``omni_lb_policy`` (that string configures omni's
-    *replica* balancer, a different layer). Key-stable (hash) request affinity is
-    not something vLLM's DP LB guarantees, so it is rejected rather than silently
-    dropped.
-    """
-    if owner != "engine":
-        _fail(
-            f"dp axis {spec.name!r} is engine data parallelism realized intra-engine by "
-            f"vLLM's DP load balancer; l1_owner must be 'engine', got {owner!r}. For "
-            "omni-coordinator-level request fan-out across independent replicas, use a "
-            "'stage_replica' axis."
-        )
-    if _is_affinity_dp_routing(spec.routing):
-        _not_implemented(
-            f"dp axis {spec.name!r} requests key-stable (hash) routing, which vLLM's "
-            "intra-engine DP load balancer does not guarantee — not supported yet. Use "
-            "RouteByStage(random|round_robin|least_queue) for stateless DP balancing."
-        )
-    if not isinstance(spec.routing, RouteByStage):
-        _fail(
-            f"dp axis {spec.name!r} expects RouteByStage(random|round_robin|least_queue) routing, "
-            f"got {type(spec.routing).__name__}"
-        )
-    if spec.routing.routing_policy not in _STAGE_POLICY_TO_OMNI_LB:
-        # Recognized-but-unimplemented routing (key-stable/hash) is already
-        # handled above via _is_affinity_dp_routing -> _not_implemented. Anything
-        # left here is an unknown/invalid policy value (e.g. a typo from YAML),
-        # i.e. invalid input -> AxisTranslationError, not NotImplementedError.
-        _fail(
-            f"dp axis {spec.name!r} has invalid routing_policy "
-            f"{spec.routing.routing_policy!r}; expected one of "
-            f"{sorted(_STAGE_POLICY_TO_OMNI_LB)}."
-        )
-
-
-def _stage_replica_lb_policy(spec: StrategySpec, owner: L1Owner) -> str:
-    """Return the omni StagePool LB policy for a delegated stage_replica axis.
-
-    A ``stage_replica`` axis is *not* a vLLM world dimension: it stands up N
-    independent engine replicas of one pipeline stage, coordinated by the omni
-    coordinator and balanced over a StagePool with stateless policies only
-    (random / round-robin / least-queue-length). It maps to the per-stage
-    ``num_replicas`` count plus the pipeline-level ``omni_lb_policy`` string, so
-    its ``l1_owner`` must be ``"delegated"`` (omni owns the routing). Key-stable
-    (hash) routing is rejected because omni has no key-stable balancer yet.
-    """
-    if owner != "delegated":
-        _fail(
-            f"stage_replica axis {spec.name!r} must be 'delegated' to omni's StagePool load "
-            f"balancer; got owner {owner!r}. Replica routing is owned by omni's coordinator."
-        )
-
-    routing = spec.routing
-    if _is_affinity_dp_routing(routing):
-        _not_implemented(
-            f"stage_replica axis {spec.name!r} requests key-stable (hash) routing, which needs a "
-            "dedicated load balancer — not implemented yet. Use "
-            "RouteByStage(random|round_robin|least_queue) to delegate to omni's load balancer."
-        )
-    if not isinstance(routing, RouteByStage):
-        _fail(
-            f"stage_replica axis {spec.name!r} expects RouteByStage(random|round_robin|least_queue) "
-            f"routing, got {type(routing).__name__}"
-        )
-    policy = _STAGE_POLICY_TO_OMNI_LB.get(routing.routing_policy)
-    if policy is None:
-        _not_implemented(
-            f"stage_replica axis {spec.name!r} routing_policy {routing.routing_policy!r} has no omni LB policy"
-        )
-    return policy
-
-
-def _validate_tp(spec: StrategySpec, owner: L1Owner) -> None:
-    if not isinstance(spec.routing, Broadcast):
-        _fail(f"tp axis {spec.name!r} expects Broadcast routing, got {type(spec.routing).__name__}")
-    if owner != "engine":
-        _fail(f"tp axis {spec.name!r} is realized intra-engine; l1_owner must be 'engine', got {owner!r}")
-
-
-def _validate_pp(spec: StrategySpec, owner: L1Owner) -> None:
-    if not isinstance(spec.routing, PipelineMicrobatch):
-        _fail(f"pp axis {spec.name!r} expects PipelineMicrobatch routing, got {type(spec.routing).__name__}")
-    if owner != "engine":
-        _fail(f"pp axis {spec.name!r} is realized intra-engine; l1_owner must be 'engine', got {owner!r}")
-
-
-def _validate_ep(spec: StrategySpec, owner: L1Owner) -> None:
-    # Dense EP: every rank still sees the whole batch, experts are sharded
-    # across ranks inside the engine (sparse MoE all-to-all is a later stage).
-    if not isinstance(spec.routing, Broadcast):
-        _fail(
-            f"ep axis {spec.name!r} expects Broadcast routing (dense expert parallel), "
-            f"got {type(spec.routing).__name__}"
-        )
-    if owner != "engine":
-        _fail(f"ep axis {spec.name!r} is realized intra-engine; l1_owner must be 'engine', got {owner!r}")
-
-
-def translate_strategy_stack(specs: Sequence[StrategySpec]) -> OmniParallelConfig:
-    """Translate a spec stack into an ``OmniParallelConfig``.
-
-    Supported kinds: dp (engine data parallel), tp, pp, (dense) ep, and
-    stage_replica (omni replicas). Raises ``NotImplementedError`` for deferred
-    (affinity / key-stable) routing, and :class:`AxisTranslationError` for any
-    other invalid spec (a kind not yet translatable, a repeated kind, an owner
-    incompatible with the axis kind, or unsupported routing). The EP degree must
-    equal tensor_parallel_size * data_parallel_size.
-    """
-    sizes: dict[str, int] = {"tensor_parallel_size": 1, "data_parallel_size": 1, "pipeline_parallel_size": 1}
-    owners: dict[MeshAxisKind, L1Owner] = {}
-    omni_lb_policy: str | None = None
-    stage_replica_size = 1
-    enable_expert_parallel = False
-    ep_size: int | None = None
-
-    for spec in specs:
-        kind = spec.mesh_axis.kind
-        if kind not in _SUPPORTED_KINDS:
-            _fail(
-                f"axis kind {kind!r} is not translatable yet (supported: {list(_SUPPORTED_KINDS)}); "
-                "it is designed-for and lands in a later stage"
-            )
-        if kind in owners:
-            _fail(f"axis kind {kind!r} appears more than once in the spec stack")
-
-        owner = _resolve_l1_owner(spec)
-        if kind == "dp":
-            _validate_dp(spec, owner)
-        elif kind == "tp":
-            _validate_tp(spec, owner)
-        elif kind == "pp":
-            _validate_pp(spec, owner)
-        elif kind == "ep":
-            _validate_ep(spec, owner)
-            enable_expert_parallel = True
-            ep_size = spec.mesh_axis.size
-        elif kind == "stage_replica":
-            omni_lb_policy = _stage_replica_lb_policy(spec, owner)
-            stage_replica_size = spec.mesh_axis.size
-
-        if kind in _AXIS_TO_ENGINE_FIELD:
-            sizes[_AXIS_TO_ENGINE_FIELD[kind]] = spec.mesh_axis.size
-        owners[kind] = owner
-
-    if enable_expert_parallel:
-        # Dense EP shards experts across exactly the TP*DP ranks, so the declared
-        # EP degree must match that product (it is not its own world dimension,
-        # and PP is excluded). An EP axis of size 1 is a degenerate no-op that
-        # still sets the flag; it is only valid when TP*DP == 1.
-        ep_ranks = sizes["tensor_parallel_size"] * sizes["data_parallel_size"]
-        if ep_size != ep_ranks:
-            _fail(
-                f"ep axis size {ep_size} must equal tensor_parallel_size*data_parallel_size "
-                f"(={ep_ranks}); dense expert parallelism shards experts across exactly those ranks"
-            )
-
-    return OmniParallelConfig(
-        tensor_parallel_size=sizes["tensor_parallel_size"],
-        data_parallel_size=sizes["data_parallel_size"],
-        pipeline_parallel_size=sizes["pipeline_parallel_size"],
-        enable_expert_parallel=enable_expert_parallel,
-        stage_replica_size=stage_replica_size,
-        omni_lb_policy=omni_lb_policy,
-        l1_owners=owners,
-    )

vllm_omni/config/config_factory.py

@@ -6,7 +6,6 @@ from __future__ import annotations
 
 import dataclasses
 import functools
-from collections.abc import Mapping
 from dataclasses import asdict
 from pathlib import Path
 from typing import Any
@@ -34,16 +33,6 @@ from vllm_omni.diffusion.utils.hf_utils import _looks_like_dreamzero
 logger = init_logger(__name__)
 
 
-# Default degree for any parallel axis / replica count that isn't set anywhere
-# (CLI, deploy YAML, or pipeline default): a single, un-parallelized rank.
-# TODO(composable_parallel): this "1" default and the per-axis device-layout
-# fallbacks are currently re-derived in the merge, apply, and reconcile layers.
-# Centralize them in one schema so the pre-spawn device guard and the engine-args
-# defaults can't drift apart. This is the light slice; the full device-layout
-# centralization is tracked as a follow-up.
-_DEFAULT_PARALLEL_DEGREE = 1
-
-
 class StageConfigFactory:
     """Factory that loads pipeline YAML and merges CLI overrides.
 
@@ -278,27 +267,14 @@ class StageConfigFactory:
         trust_remote_code: bool = False,
         cli_overrides: dict[str, Any] | None = None,
         deploy_config_path: str | None = None,
-        strategy_specs: Mapping[Any, Any] | None = None,
         **deprecated_kwargs: Any,
-    ) -> tuple[list[StageConfig] | None, str | None]:
+    ) -> list[StageConfig] | None:
         """Load pipeline + deploy config, merge with CLI overrides.
 
         Checks OMNI_PIPELINES first, since supported models should be explicitly
         registered. If a model is not registered in OMNI_PIPELINES, tries to fall
         back to using the Transformers config & finding pipelines that have overlapping
         supported architectures.
-
-        When ``strategy_specs`` is provided (a mapping of role -> list of
-        ``StrategySpec``), the derived parallel sizing is overlaid onto the
-        merged stages (see ``vllm_omni.config.composable_parallel``). This is
-        opt-in: omitting it leaves the existing merge path untouched.
-
-        Returns ``(stages, omni_lb_policy)``: the merged stages (``None`` when the
-        model is not in the pipeline registry and the caller should fall back to
-        the legacy YAML path) and the strategy-derived, pipeline-wide
-        ``omni_lb_policy`` (``None`` when no stage_replica axis set one). The
-        policy is returned rather than threaded through a mutable out-param so the
-        engine can apply it without every intermediate call carrying the dict.
         """
         if cli_overrides is None:
             cli_overrides = {}
@@ -314,9 +290,8 @@ class StageConfigFactory:
                 pipeline_cfg,
                 cli_overrides,
                 deploy_config_path,
-                strategy_specs=strategy_specs,
             )
-        return None, None
+        return None
 
     @classmethod
     def _create_from_registry(
@@ -325,17 +300,12 @@ class StageConfigFactory:
         pipeline_cfg: PipelineConfig,
         cli_overrides: dict[str, Any],
         deploy_config_path: str | None = None,
-        strategy_specs: Mapping[Any, Any] | None = None,
         **deprecated_kwargs: Any,
-    ) -> tuple[list[StageConfig], str | None]:
+    ) -> list[StageConfig]:
         """Create StageConfigs from pipeline registry + deploy YAML.
 
         Precedence: caller-typed (non-None) value > deploy YAML >
         StageDeployConfig dataclass default.
-
-        Returns ``(stages, omni_lb_policy)`` — the strategy-derived pipeline-wide
-        load-balancer policy (``None`` when no strategy set one) travels with the
-        stages instead of through a mutable out-param.
         """
         # Resolve deploy config path
         if deploy_config_path is None:
@@ -368,127 +338,12 @@ class StageConfigFactory:
 
         stages = merge_pipeline_deploy(pipeline_cfg, deploy_cfg, cli_overrides)
 
-        # Overlay declarative parallel strategies (opt-in) before CLI overrides.
-        applied = cls._apply_strategy_specs(stages, strategy_specs)
-
         explicit_overrides = {k: v for k, v in cli_overrides.items() if v is not None}
 
         for stage in stages:
             stage.runtime_overrides = cls._merge_cli_overrides(stage, explicit_overrides)
 
-        # Re-validate the resolved layout now that CLI overrides are on top.
+        return stages
-        cls._reconcile_strategy_with_cli(stages, applied)
-
-        omni_lb_policy = applied.omni_lb_policy if applied is not None else None
-        return stages, omni_lb_policy
-
-    @staticmethod
-    def _apply_strategy_specs(
-        stages: list[StageConfig],
-        strategy_specs: Mapping[Any, Any] | None,
-    ) -> Any:
-        """Overlay derived parallel sizing onto merged stages (opt-in).
-
-        ``omni_lb_policy`` cannot be set from stage configs (omni reads it once
-        at orchestrator construction), so a derived non-default policy is logged
-        here and carried on the returned ``StrategyApplyResult`` for the caller
-        to hand to the orchestrator.
-
-        Returns the ``StrategyApplyResult`` (or ``None`` when no strategy was
-        supplied) so the caller can re-validate the resolved layout once CLI
-        overrides have been merged on top, and read its ``omni_lb_policy``.
-        """
-        if not strategy_specs:
-            return None
-        from vllm_omni.config.composable_parallel import apply_strategy_specs
-
-        applied = apply_strategy_specs(stages, strategy_specs)
-        if applied.omni_lb_policy is not None:
-            logger.info(
-                "[composable_parallel] strategy derived omni_lb_policy=%r; it will be applied "
-                "to the orchestrator unless an explicit --omni-lb-policy was given.",
-                applied.omni_lb_policy,
-            )
-        return applied
-
-    @staticmethod
-    def _reconcile_strategy_with_cli(
-        stages: list[StageConfig],
-        applied: Any,
-    ) -> None:
-        """Reconcile CLI overrides applied *after* a strategy overlay.
-
-        CLI overrides are applied last (in ``to_omegaconf``), so for any stage a
-        strategy touched we must (a) warn loudly when a CLI arg overrides a
-        strategy-declared axis — the strategy is meant to be the single writer
-        for the axes it declares, so a silent CLI win is surprising — and
-        (b) re-run the device-count check against the *effective* (post-CLI)
-        ``tp``/``dp``/``pp``/``num_replicas``/``devices`` so a CLI
-        ``--devices``/``--tensor-parallel-size``/``--num-replicas`` cannot slip
-        past the pre-spawn guard that exists to prevent silent OOMs.
-        """
-        if applied is None:
-            return
-        from vllm_omni.config.composable_parallel import check_device_layout
-
-        # Axis kind -> (engine-arg field, strategy-derived attribute on the cfg).
-        axis_fields = {
-            "tp": ("tensor_parallel_size", "tensor_parallel_size"),
-            "dp": ("data_parallel_size", "data_parallel_size"),
-            "pp": ("pipeline_parallel_size", "pipeline_parallel_size"),
-        }
-
-        for stage in stages:
-            cfg = applied.per_stage_config.get(stage.stage_id)
-            if cfg is None:
-                continue
-            overrides = stage.runtime_overrides or {}
-            declared = set(cfg.l1_owners.keys())
-
-            for kind, (field_name, attr) in axis_fields.items():
-                if kind in declared and overrides.get(field_name) is not None:
-                    cli_val = overrides[field_name]
-                    derived = getattr(cfg, attr)
-                    if cli_val != derived:
-                        logger.warning(
-                            "[composable_parallel] stage %s: CLI %s=%s overrides the "
-                            "strategy-derived %s=%s. The CLI value wins; remove one to avoid ambiguity.",
-                            stage.stage_id,
-                            field_name,
-                            cli_val,
-                            field_name,
-                            derived,
-                        )
-            if "stage_replica" in declared and overrides.get("num_replicas") is not None:
-                cli_val = overrides["num_replicas"]
-                if cli_val != cfg.stage_replica_size:
-                    logger.warning(
-                        "[composable_parallel] stage %s: CLI num_replicas=%s overrides the "
-                        "strategy-derived num_replicas=%s. The CLI value wins; remove one to avoid ambiguity.",
-                        stage.stage_id,
-                        cli_val,
-                        cfg.stage_replica_size,
-                    )
-
-            def _eff(field_name: str, fallback: Any) -> Any:
-                val = overrides.get(field_name)
-                return val if val is not None else fallback
-
-            def _eff_degree(field_name: str, source: dict[str, Any]) -> int:
-                # Single place the per-axis "default to 1" fallback is applied,
-                # for both the override and the YAML-default sides (see the
-                # _DEFAULT_PARALLEL_DEGREE TODO).
-                value = _eff(field_name, source.get(field_name, _DEFAULT_PARALLEL_DEGREE))
-                return int(value or _DEFAULT_PARALLEL_DEGREE)
-
-            check_device_layout(
-                _eff("devices", stage.yaml_runtime.get("devices")),
-                tensor_parallel_size=_eff_degree("tensor_parallel_size", stage.yaml_engine_args),
-                data_parallel_size=_eff_degree("data_parallel_size", stage.yaml_engine_args),
-                pipeline_parallel_size=_eff_degree("pipeline_parallel_size", stage.yaml_engine_args),
-                num_replicas=_eff_degree("num_replicas", stage.yaml_runtime),
-                role=stage.model_stage,
-            )
 
     @classmethod
     def create_default_diffusion(cls, kwargs: dict[str, Any]) -> list[dict[str, Any]]:

vllm_omni/config/omni_config.py

@@ -551,7 +551,7 @@ class _DiffusionConfigProjection:
             elif isinstance(self.quantization_config, str):
                 self.quantization_config = build_quant_config(self.quantization_config)
             elif isinstance(self.quantization_config, Mapping):
-                self.quantization_config = dict(self.quantization_config)
+                self.quantization_config = build_quant_config(dict(self.quantization_config))
             else:
                 raise TypeError(
                     "quantization_config must be str, dict, QuantizationConfig, or None, "

vllm_omni/core/sched/omni_ar_scheduler.py

@@ -314,6 +314,13 @@ class OmniARScheduler(OmniSchedulerMixin, VLLMScheduler):
 
         outputs: dict[int, list[EngineCoreOutput]] = defaultdict(list)
         spec_decoding_stats: SpecDecodingStats | None = None
+        kv_connector_stats: KVConnectorStats | None = (
+            kv_connector_output.kv_connector_stats if kv_connector_output else None
+        )
+        if kv_connector_stats and self.connector:
+            kv_stats = self.connector.get_kv_connector_stats()
+            if kv_stats:
+                kv_connector_stats = kv_connector_stats.aggregate(kv_stats)
 
         failed_kv_load_req_ids = None
         if kv_connector_output and kv_connector_output.invalid_block_ids:
@@ -556,20 +563,6 @@ class OmniARScheduler(OmniSchedulerMixin, VLLMScheduler):
         if kv_connector_output:
             self._update_from_kv_xfer_finished(kv_connector_output)
 
-        # Worker-side KV connector stats from the model runner output.
-        kv_connector_stats: KVConnectorStats | None = (
-            kv_connector_output.kv_connector_stats if kv_connector_output else None
-        )
-        if self.connector:
-            # Scheduler-side KV connector stats collected after connector update.
-            scheduler_kv_connector_stats = self.connector.get_kv_connector_stats()
-            if scheduler_kv_connector_stats is not None and not scheduler_kv_connector_stats.is_empty():
-                kv_connector_stats = (
-                    kv_connector_stats.aggregate(scheduler_kv_connector_stats)
-                    if kv_connector_stats is not None
-                    else scheduler_kv_connector_stats
-                )
-
         # collect KV cache events from KV cache manager
         events = self.kv_cache_manager.take_events()
 

vllm_omni/core/sched/omni_generation_scheduler.py

@@ -431,6 +431,14 @@ class OmniGenerationScheduler(OmniSchedulerMixin, VLLMScheduler):
 
         outputs: dict[int, list[EngineCoreOutput]] = defaultdict(list)
         spec_decoding_stats: SpecDecodingStats | None = None
+        kv_connector_stats: KVConnectorStats | None = (
+            kv_connector_output.kv_connector_stats if kv_connector_output else None
+        )
+        # Merge connector-side stats (align with v0.14.0)
+        if kv_connector_stats and self.connector:
+            kv_stats = self.connector.get_kv_connector_stats()
+            if kv_stats:
+                kv_connector_stats = kv_connector_stats.aggregate(kv_stats)
 
         failed_kv_load_req_ids = None
         if kv_connector_output and getattr(kv_connector_output, "invalid_block_ids", None):
@@ -638,20 +646,6 @@ class OmniGenerationScheduler(OmniSchedulerMixin, VLLMScheduler):
         if kv_connector_output:
             self._update_from_kv_xfer_finished(kv_connector_output)
 
-        # Worker-side KV connector stats from the model runner output.
-        kv_connector_stats: KVConnectorStats | None = (
-            kv_connector_output.kv_connector_stats if kv_connector_output else None
-        )
-        if self.connector:
-            # Scheduler-side KV connector stats collected after connector update.
-            scheduler_kv_connector_stats = self.connector.get_kv_connector_stats()
-            if scheduler_kv_connector_stats is not None and not scheduler_kv_connector_stats.is_empty():
-                kv_connector_stats = (
-                    kv_connector_stats.aggregate(scheduler_kv_connector_stats)
-                    if kv_connector_stats is not None
-                    else scheduler_kv_connector_stats
-                )
-
         # Collect and publish KV cache events (align with v0.14.0)
         events = self.kv_cache_manager.take_events()
         if self.connector is not None:

vllm_omni/deploy/ming_tts.yaml

@@ -15,7 +15,6 @@ connectors:
     name: SharedMemoryConnector
     extra:
       latent_chunk_size: 25
-      initial_latent_chunk_size: 4
       latent_left_context: 0
 
 stages:

vllm_omni/deploy/ming_tts_moe.yaml

@@ -23,7 +23,6 @@ connectors:
     name: SharedMemoryConnector
     extra:
       latent_chunk_size: 25
-      initial_latent_chunk_size: 4
       latent_left_context: 0
 
 stages:

vllm_omni/deploy/qwen3_omni_moe.yaml

@@ -47,6 +47,7 @@ stages:
       temperature: 0.9
       top_k: 50
       max_tokens: 4096
+      seed: 42
       repetition_penalty: 1.05
 
   - stage_id: 2
@@ -66,6 +67,7 @@ stages:
       top_p: 1.0
       top_k: -1
       max_tokens: 65536
+      seed: 42
       repetition_penalty: 1.1
 
 platforms: