[Bugfix] Fix CLI defaults overriding stage configs

benchmarks/diffusion/quantization_quality.py

@@ -58,6 +58,8 @@
 import numpy as np
 import torch
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
+
 
 def compute_lpips_images(
     baseline_images: list,
@@ -137,6 +139,7 @@ def _build_omni_kwargs(args, quantization=None):
     )
     kwargs = {
         "model": args.model,
+        "explicit_overrides": getattr(args, "_explicit_overrides", None),
         "parallel_config": parallel_config,
         "enforce_eager": args.enforce_eager,
     }
@@ -452,7 +455,7 @@ def parse_args():
     parser.add_argument("--ring-degree", type=int, default=1)
     parser.add_argument("--tensor-parallel-size", type=int, default=1)
     parser.add_argument("--enforce-eager", action="store_true")
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 if __name__ == "__main__":

examples/offline_inference/bagel/end2end.py

@@ -1,6 +1,7 @@
 import argparse
 import os
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.inputs.data import OmniPromptType
 from vllm_omni.model_executor.stage_input_processors.bagel import (
     GEN_THINK_SYSTEM_PROMPT,
@@ -98,7 +99,7 @@ def parse_args():
         help="Enable thinking mode: AR stage decodes <think>...</think> planning tokens before image generation.",
     )
 
-    args = parser.parse_args()
+    args = parse_args_with_explicit_overrides(parser)
     return args
 
 
@@ -152,6 +153,7 @@ def main():
     )
     if args.quantization:
         omni_kwargs["quantization_config"] = args.quantization
+    omni_kwargs["explicit_overrides"] = getattr(args, "_explicit_overrides", None)
 
     omni = Omni(model=model_name, **omni_kwargs)
 

examples/offline_inference/cosyvoice3/verify_e2e_cosyvoice.py

@@ -10,6 +10,7 @@
 from vllm import SamplingParams
 from vllm.assets.audio import AudioAsset
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.model_executor.models.cosyvoice3.config import CosyVoice3Config
 from vllm_omni.model_executor.models.cosyvoice3.tokenizer import get_qwen_tokenizer
@@ -55,7 +56,7 @@ def run_e2e():
         required=True,
         help="Path to tokenizer directory (e.g., <model_path>/CosyVoice-BlankEN).",
     )
-    args = parser.parse_args()
+    args = parse_args_with_explicit_overrides(parser)
     _ensure_mel_filters_asset()
     # Ensure tokenizer directory exists
     if not os.path.exists(args.tokenizer):
@@ -72,6 +73,7 @@ def run_e2e():
     # We pass trust_remote_code=True same as Qwen examples
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         stage_configs_path=args.stage_config,
         trust_remote_code=True,
         tokenizer=args.tokenizer,

examples/offline_inference/custom_pipeline/image_to_image/image_edit.py

@@ -52,6 +52,7 @@
 from PIL import Image
 
 from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 from vllm_omni.outputs import OmniRequestOutput
@@ -100,7 +101,7 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--vae-use-tiling", action="store_true")
     parser.add_argument("--enable-cpu-offload", action="store_true")
 
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 # ===========================
@@ -149,6 +150,7 @@ async def main():
     # ---- Initialize Omni ----
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         vae_use_slicing=args.vae_use_slicing,
         vae_use_tiling=args.vae_use_tiling,
         cache_backend=args.cache_backend,

examples/offline_inference/dynin_omni/end2end.py

@@ -18,6 +18,8 @@
 import torch
 from PIL import Image
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
+
 TASK_CHOICES = ("t2t", "t2i", "t2s", "i2i", "i2t", "s2t", "v2t")
 
 TASK_DEFAULT_RUNTIME = {
@@ -970,7 +972,7 @@ def parse_args(repo_root: Path) -> argparse.Namespace:
     parser.add_argument("--vq-model-audio-local-files-only", action=argparse.BooleanOptionalAction, default=None)
 
     parser.add_argument("--disable-hf-xet", action=argparse.BooleanOptionalAction, default=True)
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def main() -> None:
@@ -1395,7 +1397,12 @@ def main() -> None:
     from vllm_omni.entrypoints.omni import Omni
 
     stage_config_path = str(Path(args.stage_config_path).expanduser())
-    omni = Omni(model=model_source, stage_configs_path=stage_config_path, dtype=args.dtype)
+    omni = Omni(
+        model=model_source,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
+        stage_configs_path=stage_config_path,
+        dtype=args.dtype,
+    )
     sampling_params_list = [
         SamplingParams(max_tokens=int(args.max_tokens_per_stage), temperature=0.0, top_p=1.0, detokenize=False)
         for _ in range(omni.num_stages)

examples/offline_inference/fish_speech/end2end.py

@@ -29,6 +29,7 @@
 
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni import AsyncOmni, Omni
 from vllm_omni.model_executor.models.fish_speech.dac_utils import DAC_HOP_LENGTH, DAC_SAMPLE_RATE
 from vllm_omni.model_executor.models.fish_speech.prompt_utils import (
@@ -149,6 +150,7 @@ def main(args):
 
     omni = Omni(
         model=model_name,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         stage_configs_path=stage_configs_path,
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
@@ -185,6 +187,7 @@ async def main_streaming(args):
 
     omni = AsyncOmni(
         model=model_name,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         stage_configs_path=stage_configs_path,
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
@@ -273,7 +276,7 @@ def parse_args():
         default=False,
         help="Stream audio chunks as they arrive via AsyncOmni.",
     )
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 if __name__ == "__main__":

examples/offline_inference/glm_image/end2end.py

@@ -44,6 +44,7 @@
 
 from PIL import Image
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 
@@ -260,6 +261,7 @@ def main(args: argparse.Namespace) -> None:
 
     omni = Omni(
         model=args.model_path,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         stage_configs_path=config_path,
         log_stats=args.enable_stats,
         stage_init_timeout=args.stage_init_timeout,
@@ -503,7 +505,7 @@ def parse_args() -> argparse.Namespace:
         help="Enable diffusion pipeline profiler to display stage durations.",
     )
 
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 if __name__ == "__main__":

examples/offline_inference/helios/end2end.py

@@ -52,6 +52,7 @@
 import torch
 
 from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 from vllm_omni.outputs import OmniRequestOutput
@@ -196,7 +197,7 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--cfg-parallel-size", type=int, default=1, choices=[1, 2], help="CFG parallel size.")
     parser.add_argument("--tensor-parallel-size", type=int, default=1, help="Tensor parallelism size.")
 
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def main():
@@ -212,6 +213,7 @@ def main():
 
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         enable_layerwise_offload=args.enable_layerwise_offload,
         vae_use_slicing=args.vae_use_slicing,
         vae_use_tiling=args.vae_use_tiling,

examples/offline_inference/hunyuan_image3/image_to_text.py

@@ -6,6 +6,7 @@
 
 from PIL import Image
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 
 """
@@ -46,7 +47,7 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable diffusion pipeline profiler to display stage durations.",
     )
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def load_image(image_path: str) -> Image.Image:
@@ -59,6 +60,7 @@ def load_image(image_path: str) -> Image.Image:
 def main(args: argparse.Namespace) -> None:
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,
         mode="image-to-text",
     )

examples/offline_inference/image_to_image/image_edit.py

@@ -95,6 +95,7 @@
 from PIL import Image
 
 from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 from vllm_omni.outputs import OmniRequestOutput
@@ -330,7 +331,7 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable diffusion pipeline profiler to display stage durations.",
     )
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def main():
@@ -386,6 +387,7 @@ def main():
     # Initialize Omni with appropriate pipeline
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         enable_layerwise_offload=args.enable_layerwise_offload,
         vae_use_slicing=args.vae_use_slicing,
         vae_use_tiling=args.vae_use_tiling,

examples/offline_inference/image_to_video/image_to_video.py

@@ -42,6 +42,7 @@
 import torch
 
 from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 from vllm_omni.outputs import OmniRequestOutput
@@ -187,7 +188,7 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable diffusion pipeline profiler to display stage durations.",
     )
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def calculate_dimensions(
@@ -281,6 +282,7 @@ def main():
     )
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         enable_layerwise_offload=args.enable_layerwise_offload,
         vae_use_slicing=args.vae_use_slicing,
         vae_use_tiling=args.vae_use_tiling,

examples/offline_inference/magi_human/end2end.py

@@ -1,6 +1,7 @@
 import argparse
 
 from vllm_omni.diffusion.utils.media_utils import mux_video_audio_bytes
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 
@@ -24,7 +25,7 @@ def parse_args():
     parser.add_argument("--width", type=int, default=448, help="Video width.")
     parser.add_argument("--num-inference-steps", type=int, default=8, help="Number of denoising steps.")
     parser.add_argument("--seed", type=int, default=52, help="Random seed for generation.")
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def main():
@@ -33,6 +34,7 @@ def main():
     print(f"Initializing MagiHuman pipeline with TP={args.tensor_parallel_size}...")
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         init_timeout=1200,
         tensor_parallel_size=args.tensor_parallel_size,
         devices=list(range(args.tensor_parallel_size)),

examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py

@@ -19,6 +19,7 @@
 from vllm.multimodal.image import convert_image_mode
 
 from vllm_omni import Omni
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 
 DEFAULT_SYSTEM = "You are a helpful assistant."
 DEFAULT_QUESTION = "Please summarize the content of this image."
@@ -48,7 +49,7 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable diffusion pipeline profiler to display stage durations.",
     )
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 def build_prompt(system: str, question: str) -> str:
@@ -73,6 +74,7 @@ def main() -> None:
 
     omni = Omni(
         model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         stage_configs_path=args.stage_config,
         trust_remote_code=args.trust_remote_code,
         enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,

examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_t2i.py

@@ -29,6 +29,7 @@
 from vllm.sampling_params import SamplingParams
 
 from vllm_omni import Omni
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logger = logging.getLogger(__name__)
@@ -117,7 +118,7 @@ def parse_args() -> argparse.Namespace:
     )
     p.add_argument("--out", type=str, default="output.png", help="Path to save the generated image.")
     p.add_argument("--trust-remote-code", action="store_true", help="Trust remote code when loading the model.")
-    args = p.parse_args()
+    args = parse_args_with_explicit_overrides(p)
     if not args.prompt:
         args.prompt = ["A stylish woman with sunglasses riding a motorcycle in NYC."]
     return args
@@ -194,7 +195,12 @@ def main() -> None:
     expected_grid_tokens = ar_height * (ar_width + 1)
 
     logger.info("Initializing Omni pipeline...")
-    omni = Omni(model=args.model, stage_configs_path=args.stage_config, trust_remote_code=args.trust_remote_code)
+    omni = Omni(
+        model=args.model,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
+        stage_configs_path=args.stage_config,
+        trust_remote_code=args.trust_remote_code,
+    )
     try:
         ar_sampling = SamplingParams(
             temperature=1.0,

examples/offline_inference/mimo_audio/end2end.py

@@ -23,6 +23,7 @@
 from vllm import SamplingParams
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
+from vllm_omni.entrypoints.utils import parse_args_with_explicit_overrides
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniTokensPrompt
 
@@ -182,6 +183,7 @@ def main(args):
 
     omni = Omni(
         model=model_name,
+        explicit_overrides=getattr(args, "_explicit_overrides", None),
         stage_configs_path=args.stage_configs_path,
         log_stats=args.enable_stats,
         log_file=("omni_pipeline.log" if args.enable_stats else None),
@@ -434,7 +436,7 @@ def parse_args():
         help="Path to a stage configs file.",
     )
 
-    return parser.parse_args()
+    return parse_args_with_explicit_overrides(parser)
 
 
 if __name__ == "__main__":