[Refactor][OutputProcessor 2/8]: OmniRequestOutput should inherit RequestOutput and no nested wrap-up (#5146)

Signed-off-by: Boao Shi <aoibosh@connect.hku.hk>
Signed-off-by: bowieshi <aoibosh@connect.hku.hk>
Signed-off-by: princepride <wangzhipeng628@gmail.com>
Signed-off-by: R2-Y <ruiruyang2@gmail.com>
Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
Signed-off-by: Shi Boao Bowie <aoibosh@connect.hku.hk>
Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com>
Co-authored-by: ruirui(rein) yang <73573651+R2-Y@users.noreply.github.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

examples/offline_inference/aura_omni/end2end.py

@@ -131,7 +131,7 @@ def build_sampling_params() -> list[SamplingParams]:
 
 
 def save_stage_output(stage_outputs, output_dir: str) -> None:
-    output = stage_outputs.request_output
+    output = stage_outputs
     request_id = output.request_id
     if stage_outputs.final_output_type == "text":
         text_output = output.outputs[0].text

examples/offline_inference/covo_audio/end2end.py

@@ -104,7 +104,7 @@ def main(args):
     os.makedirs(output_dir, exist_ok=True)
 
     for stage_outputs in omni_outputs:
-        output = stage_outputs.request_output
+        output = stage_outputs
         if stage_outputs.final_output_type == "text":
             request_id = output.request_id
             text_output = output.outputs[0].text

examples/offline_inference/custom_pipeline/image_to_image/image_edit.py

@@ -228,7 +228,7 @@ async def main():
     if not outputs:
         raise ValueError("No output produced from omni.generate()")
 
-    first_out = outputs[0].request_output
+    first_out = outputs[0]
     req_out: OmniRequestOutput = first_out
 
     # Verify trajectory data (from custom pipeline)

examples/offline_inference/dynin_omni/end2end.py

@@ -741,7 +741,7 @@ def make_mmu_prompt(
 
 def iter_mm_outputs(outputs: list[Any]):
     for omni_out in outputs:
-        req_out = getattr(omni_out, "request_output", None)
+        req_out = omni_out
         req_list = req_out if isinstance(req_out, list) else [req_out]
         for item in req_list:
             if item is None:

examples/offline_inference/hunyuan_image3/end2end.py

@@ -303,7 +303,7 @@ def main():
     )
     img_idx = 0
     for req_output in omni_outputs:
-        ro = getattr(req_output, "request_output", None)
+        ro = req_output
         stage_id = getattr(req_output, "stage_id", None)
 
         # AR stage text — each CompletionOutput.text is already a delta when

examples/offline_inference/image_to_image/image_edit.py

@@ -621,7 +621,7 @@ def main():
         images = getattr(output, "images", None)
         if images:
             break
-        req_out = getattr(output, "request_output", None)
+        req_out = output
         images = getattr(req_out, "images", None) if req_out is not None else None
         if images:
             break

examples/offline_inference/image_to_video/image_to_video.py

@@ -646,8 +646,8 @@ def main():
         if frames.multimodal_output and "audio" in frames.multimodal_output:
             audio = frames.multimodal_output["audio"]
             audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
-        if frames.is_pipeline_output and frames.request_output is not None:
-            inner_output = frames.request_output
+        if frames.is_pipeline_output and frames is not None:
+            inner_output = frames
             if isinstance(inner_output, OmniRequestOutput):
                 if inner_output.multimodal_output and "audio" in inner_output.multimodal_output:
                     audio = inner_output.multimodal_output["audio"]

examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py

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

examples/offline_inference/mimo_audio/end2end.py

@@ -297,7 +297,7 @@ def main(args):
     os.makedirs(output_dir, exist_ok=True)
 
     for stage_outputs in omni_outputs:
-        output = stage_outputs.request_output
+        output = stage_outputs
         if stage_outputs.final_output_type == "text":
             request_id = output.request_id
             text_output = output.outputs[0].text

examples/offline_inference/ming_flash_omni/end2end.py

@@ -347,7 +347,7 @@ def main(args):
         omni.start_profile(stages=args.profiler_stages)
 
     for stage_outputs in omni.generate(prompts, sampling_params_list):
-        output = stage_outputs.request_output
+        output = stage_outputs
         if stage_outputs.final_output_type == "text":
             request_id = output.request_id
             text_output = output.outputs[0].text

examples/offline_inference/qwen2_5_omni/end2end.py

@@ -392,7 +392,7 @@ def main(args):
     total_requests = len(prompts)
     processed_count = 0
     for stage_outputs in omni_generator:
-        output = stage_outputs.request_output
+        output = stage_outputs
         if stage_outputs.final_output_type == "text":
             request_id = output.request_id
             text_output = output.outputs[0].text

examples/offline_inference/qwen3_omni/end2end.py

@@ -367,7 +367,7 @@ def main(args):
     print(f"query type: {args.query_type}")
 
     for stage_outputs in omni_generator:
-        output = stage_outputs.request_output
+        output = stage_outputs
         if stage_outputs.final_output_type == "text":
             request_id = output.request_id
             text_output = output.outputs[0].text

examples/offline_inference/qwen3_omni/end2end_async_chunk.py

@@ -237,7 +237,7 @@ async def run_single_request(
             sampling_params_list=sampling_params_list,
             output_modalities=output_modalities,
         ):
-            output = omni_output.request_output
+            output = omni_output
             if omni_output.final_output_type == "text":
                 if stage_0_first_output_ts is None:
                     stage_0_first_output_ts = time.perf_counter()

examples/offline_inference/speech_to_video/speech_to_video.py

@@ -26,7 +26,6 @@ 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
 
 
@@ -352,8 +351,8 @@ def main():
         else:
             print("[Profiler] No valid profiling data returned.")
 
-    # Extract output from result
-    output = OmniRequestOutput.unwrap_result(result)
+    # omni.generate() returns a list for sync calls; unwrap single-result list.
+    output = result[0] if isinstance(result, list) else result
 
     if not output.images:
         raise ValueError("No video frames found in OmniRequestOutput.")

examples/offline_inference/step_audio2/end2end.py

@@ -345,7 +345,7 @@ def main(args):
     # Process outputs from each stage
     for stage_idx, stage_outputs in enumerate(omni_outputs):
         # request_output may be a single object or a list
-        raw_output = stage_outputs.request_output
+        raw_output = stage_outputs
         outputs_list = raw_output if isinstance(raw_output, list) else [raw_output]
 
         if stage_outputs.final_output_type == "text":

examples/offline_inference/text_to_audio/text_to_audio.py

@@ -43,6 +43,7 @@ from vllm_omni.diffusion.utils.param_utils import apply_declared_extra_args
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 from vllm_omni.model_extras import get_extra_body_params, get_model_class_name
+from vllm_omni.outputs import OmniRequestOutput
 from vllm_omni.platforms import current_omni_platform
 
 AUDIOX_TASKS = ("t2a", "t2m", "v2a", "v2m", "tv2a", "tv2m")
@@ -388,9 +389,9 @@ def main():
         raise ValueError("No output generated from omni.generate()")
 
     output = outputs[0]
-    if not hasattr(output, "request_output") or not output.request_output:
+    if not isinstance(output, OmniRequestOutput) or not output:
         raise ValueError("No request_output found in OmniRequestOutput")
-    request_output = output.request_output
+    request_output = output
     if not hasattr(request_output, "multimodal_output"):
         raise ValueError("No multimodal_output found in request_output")
 

examples/offline_inference/text_to_image/gradio_demo.py

@@ -115,7 +115,7 @@ def build_demo(args: argparse.Namespace) -> gr.Blocks:
         )
         images_outputs = []
         for output in outputs:
-            req_out = output.request_output
+            req_out = output
             if not isinstance(req_out, OmniRequestOutput) or not hasattr(req_out, "images"):
                 raise ValueError("Invalid request_output structure or missing 'images' key")
             images = req_out.images

examples/offline_inference/text_to_image/text_to_image.py

@@ -656,7 +656,7 @@ def main():
         images = getattr(output, "images", None)
         if images:
             break
-        req_out = getattr(output, "request_output", None)
+        req_out = output
         images = getattr(req_out, "images", None) if req_out is not None else None
         if images:
             break

examples/offline_inference/text_to_speech/cosyvoice3/end2end.py

@@ -168,7 +168,7 @@ def run_e2e():
     print(f"Received {len(outputs)} outputs.")
     for i, output in enumerate(outputs):
         try:
-            ro = output.request_output
+            ro = output
             if ro is None:
                 print("No request_output found.")
                 continue

examples/offline_inference/text_to_speech/fish_speech/end2end.py

@@ -179,7 +179,7 @@ def main(args):
 
     t_start = time.perf_counter()
     for stage_outputs in omni.generate(inputs):
-        request_output = stage_outputs.request_output
+        request_output = stage_outputs
         if request_output is None or not request_output.outputs:
             continue
         _save_wav(
@@ -215,7 +215,7 @@ async def main_streaming(args):
     sample_rate = None
 
     async for stage_output in omni.generate(prompt, request_id=request_id):
-        mm = stage_output.request_output.outputs[0].multimodal_output
+        mm = stage_output.outputs[0].multimodal_output
         if not stage_output.finished:
             t_now = time.perf_counter()
             sample_rate, n = _collect_audio_chunk(mm, all_audio_chunks, sample_rate)

examples/offline_inference/text_to_speech/moss_tts_nano/end2end.py

@@ -137,7 +137,7 @@ def main(args) -> None:
     params_list = sampling_params
 
     for stage_outputs in omni.generate(inputs, params_list):
-        for i, req_output in enumerate(stage_outputs.request_output):
+        for i, req_output in enumerate(stage_outputs):
             for j, out in enumerate(req_output.outputs):
                 mm = out.multimodal_output
                 if mm is None:

examples/offline_inference/text_to_speech/omnivoice/end2end.py

@@ -135,7 +135,7 @@ def run_e2e():
     print(f"Received {len(outputs)} outputs.")
     for i, output in enumerate(outputs):
         try:
-            ro = output.request_output
+            ro = output
             if ro is None:
                 print("No request_output found.")
                 continue

examples/offline_inference/text_to_speech/qwen3_tts/end2end.py

@@ -391,7 +391,7 @@ def main(args):
     for batch_start in range(0, len(inputs), batch_size):
         batch = inputs[batch_start : batch_start + batch_size]
         for stage_outputs in omni.generate(batch):
-            output = stage_outputs.request_output
+            output = stage_outputs
             _save_wav(output_dir, output.request_id, output.outputs[0].multimodal_output)
 
 
@@ -412,7 +412,7 @@ async def main_streaming(args):
         t_prev = t_start
         chunk_idx = 0
         async for stage_output in omni.generate(prompt, request_id=request_id):
-            mm = stage_output.request_output.outputs[0].multimodal_output
+            mm = stage_output.outputs[0].multimodal_output
             if not stage_output.finished:
                 t_now = time.perf_counter()
                 audio = mm.get("audio")

examples/offline_inference/text_to_speech/qwen3_tts/word_timestamps.py

@@ -0,0 +1,172 @@
+"""Offline Qwen3-TTS example with word-level timestamps.
+
+This demo runs Qwen3-TTS offline, aligns the synthesized audio in memory with
+the same shared forced aligner used by the streaming server path, then writes
+the WAV and JSON sidecar:
+
+    python examples/offline_inference/text_to_speech/qwen3_tts/word_timestamps.py \
+        --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+        --forced-aligner Qwen/Qwen3-ForcedAligner-0.6B
+
+On machines without a local CUDA toolkit, set ``VLLM_USE_FLASHINFER_SAMPLER=0``
+to avoid FlashInfer sampler JIT during warmup.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import os
+from pathlib import Path
+from typing import Any
+
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+# The in-process offline aligner consumes token_classify outputs. Keeping V1
+# multiprocessing off avoids msgspec serialization issues for token-wise logits.
+os.environ.setdefault("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
+
+import numpy as np
+import soundfile as sf
+import torch
+from end2end import _estimate_prompt_len
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+from vllm_omni import Omni
+from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
+from vllm_omni.utils.forced_aligner import align, build_forced_aligner_config
+
+
+def _default_stage_config() -> str:
+    repo_root = Path(__file__).resolve().parents[4]
+    return str(repo_root / "vllm_omni" / "deploy" / "qwen3_tts.yaml")
+
+
+def _build_custom_voice_input(args: Any) -> dict[str, Any]:
+    additional_information = {
+        "task_type": ["CustomVoice"],
+        "text": [args.text],
+        "language": [args.language],
+        "speaker": [args.speaker],
+        "instruct": [args.instructions],
+        "max_new_tokens": [args.max_new_tokens],
+    }
+    return {
+        "prompt_token_ids": [0] * _estimate_prompt_len(additional_information, args.model),
+        "additional_information": additional_information,
+    }
+
+
+def _audio_tensor_and_sample_rate(mm: dict[str, Any]) -> tuple[torch.Tensor, int]:
+    audio_data = mm["audio"]
+    sr_raw = mm["sr"]
+    sr_val = sr_raw[-1] if isinstance(sr_raw, list) and sr_raw else sr_raw
+    sample_rate = int(sr_val.item()) if hasattr(sr_val, "item") else int(sr_val)
+    audio_tensor = torch.cat(audio_data, dim=-1) if isinstance(audio_data, list) else audio_data
+    return audio_tensor.float().detach().cpu().flatten(), sample_rate
+
+
+def _float_audio_to_pcm16_bytes(audio: torch.Tensor) -> bytes:
+    samples = audio.numpy()
+    pcm = (np.clip(samples, -1.0, 1.0) * 32767.0).astype("<i2")
+    return pcm.tobytes()
+
+
+async def _run_alignment(args: Any, audio: torch.Tensor, sample_rate: int) -> list[dict[str, Any]] | None:
+    aligner_config = build_forced_aligner_config(args)
+    if aligner_config is None:
+        raise ValueError("--forced-aligner or --forced-aligner-config is required")
+    timestamps = await align(
+        audio=_float_audio_to_pcm16_bytes(audio),
+        text=args.text,
+        sample_rate=sample_rate,
+        config=aligner_config,
+        language=args.language,
+    )
+    if timestamps is None:
+        return None
+    return [
+        {
+            "word": item.word,
+            "start_ms": item.start_ms,
+            "end_ms": item.end_ms,
+        }
+        for item in timestamps
+    ]
+
+
+def main(args: Any) -> None:
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    omni_kwargs = vars(args).copy()
+    for key in (
+        "forced_aligner",
+        "text",
+        "language",
+        "speaker",
+        "instructions",
+        "max_new_tokens",
+        "output_dir",
+    ):
+        omni_kwargs.pop(key, None)
+    omni_kwargs["stage_configs_path"] = args.stage_configs_path or _default_stage_config()
+    omni_kwargs["log_stats"] = args.log_stats
+    omni = Omni(**omni_kwargs)
+
+    prompt = _build_custom_voice_input(args)
+    final_output = None
+    for stage_outputs in omni.generate([prompt]):
+        final_output = stage_outputs
+    if final_output is None:
+        raise RuntimeError("Qwen3-TTS did not produce an output.")
+
+    mm = final_output.outputs[0].multimodal_output
+    audio, sample_rate = _audio_tensor_and_sample_rate(mm)
+
+    wav_path = output_dir / "qwen3_tts_word_timestamps.wav"
+    # Align the in-memory tensor, not the saved WAV. The WAV is only a demo
+    # artifact so users can listen to the same audio referenced by the sidecar.
+    timestamps = asyncio.run(_run_alignment(args, audio, sample_rate))
+    sf.write(wav_path, audio.numpy(), samplerate=sample_rate, format="WAV")
+
+    sidecar = {
+        "text": args.text,
+        "sample_rate": sample_rate,
+        "audio_path": str(wav_path),
+        "timestamps": timestamps,
+    }
+    json_path = output_dir / "qwen3_tts_word_timestamps.json"
+    json_path.write_text(json.dumps(sidecar, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+
+    print(f"Saved audio: {wav_path}")
+    print(f"Saved timestamps: {json_path}")
+    print(json.dumps(sidecar["timestamps"], ensure_ascii=False, indent=2))
+
+
+def parse_args() -> Any:
+    parser = FlexibleArgumentParser(description="Offline Qwen3-TTS word timestamp example")
+    parser.add_argument("--model", default="Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice", help="Qwen3-TTS model path/name")
+    parser.add_argument(
+        "--forced-aligner",
+        default=None,
+        help="Qwen3 forced aligner model path/name",
+    )
+    parser.add_argument(
+        "--forced-aligner-config",
+        default=None,
+        help="Optional YAML file for forced aligner settings (incl. gpu_memory_utilization)",
+    )
+    parser.add_argument("--stage-configs-path", default=None, help="Qwen3-TTS deploy YAML")
+    parser.add_argument("--text", default="Hello world.", help="Text to synthesize and align")
+    parser.add_argument("--language", default="English", help="Qwen3-TTS language field")
+    parser.add_argument("--speaker", default="Vivian", help="CustomVoice speaker name")
+    parser.add_argument("--instructions", default="", help="Optional speaking style instruction")
+    parser.add_argument("--max-new-tokens", type=int, default=2048, help="TTS max_new_tokens")
+    parser.add_argument("--output-dir", default="output_audio", help="Directory for WAV and JSON sidecar")
+    parser.add_argument("--log-stats", action="store_true", default=False, help="Enable vLLM-Omni stats logging")
+    nullify_stage_engine_defaults(parser)
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    main(parse_args())

examples/offline_inference/text_to_speech/soulxsinger/benchmark.py

@@ -38,7 +38,7 @@ def _stage_durations_from_output(outputs) -> dict[str, float]:
     durations = getattr(omni_out, "stage_durations", None) or {}
     if durations:
         return dict(durations)
-    ro = getattr(omni_out, "request_output", None)
+    ro = omni_out
     if ro is not None:
         return dict(getattr(ro, "stage_durations", None) or {})
     return {}

examples/offline_inference/text_to_speech/soulxsinger/end2end.py

@@ -99,7 +99,7 @@ def extract_audio(outputs) -> tuple[np.ndarray, int]:
     out = outputs[0]
     mm = getattr(out, "multimodal_output", None)
     if mm is None:
-        ro = getattr(out, "request_output", None)
+        ro = out
         if ro is not None:
             mm = getattr(ro, "multimodal_output", None)
             if mm is None and ro.outputs:

examples/offline_inference/text_to_video/text_to_video.py

@@ -390,7 +390,7 @@ def _extract_peak_memory_mb(result: Any) -> float:
         return 0.0
     val = getattr(result, "peak_memory_mb", 0.0)
     if not val:
-        inner = getattr(result, "request_output", None)
+        inner = result
         if isinstance(inner, list):
             inner = inner[0] if inner else None
         val = getattr(inner, "peak_memory_mb", 0.0)
@@ -575,8 +575,8 @@ def main():
         if frames.multimodal_output and "audio" in frames.multimodal_output:
             audio = frames.multimodal_output["audio"]
             audio_sample_rate = frames.multimodal_output.get("audio_sample_rate", audio_sample_rate)
-        if frames.is_pipeline_output and frames.request_output is not None:
-            inner_output = frames.request_output
+        if frames.is_pipeline_output and frames is not None:
+            inner_output = frames
             if isinstance(inner_output, OmniRequestOutput):
                 if inner_output.multimodal_output and "audio" in inner_output.multimodal_output:
                     audio = inner_output.multimodal_output["audio"]

examples/offline_inference/x_to_video_audio/x_to_video_audio.py

@@ -138,7 +138,7 @@ def _extract_peak_memory_mb(result: Any) -> float:
         return 0.0
     val = getattr(result, "peak_memory_mb", 0.0)
     if not val:
-        inner = getattr(result, "request_output", None)
+        inner = result
         if isinstance(inner, list):
             inner = inner[0] if inner else None
         val = getattr(inner, "peak_memory_mb", 0.0)