model: Nemotron-omni-v3-alias (#23857)

python/sglang/srt/configs/__init__.py

@@ -19,7 +19,10 @@ from sglang.srt.configs.lfm2 import Lfm2Config
 from sglang.srt.configs.lfm2_moe import Lfm2MoeConfig
 from sglang.srt.configs.lfm2_vl import Lfm2VlConfig
 from sglang.srt.configs.longcat_flash import LongcatFlashConfig
-from sglang.srt.configs.nano_nemotron_vl import NemotronH_Nano_VL_V2_Config
+from sglang.srt.configs.nano_nemotron_vl import (
+    NemotronH_Nano_Omni_Reasoning_V3_Config,
+    NemotronH_Nano_VL_V2_Config,
+)
 from sglang.srt.configs.nemotron_h import NemotronHConfig
 from sglang.srt.configs.olmo3 import Olmo3Config
 from sglang.srt.configs.qwen3_5 import Qwen3_5Config, Qwen3_5MoeConfig
@@ -61,6 +64,7 @@ __all__ = [
     "Lfm2VlConfig",
     "NemotronHConfig",
     "NemotronH_Nano_VL_V2_Config",
+    "NemotronH_Nano_Omni_Reasoning_V3_Config",
     "JetNemotronConfig",
     "JetVLMConfig",
     "Step3p5Config",

python/sglang/srt/configs/model_config.py

@@ -1387,6 +1387,7 @@ multimodal_model_archs = [
     "MllamaForConditionalGeneration",
     "MossVLForConditionalGeneration",
     "NemotronH_Nano_VL_V2",
+    "NemotronH_Nano_Omni_Reasoning_V3",
     "PixtralForConditionalGeneration",
     "Qwen2AudioForConditionalGeneration",
     "Qwen2VLForConditionalGeneration",

python/sglang/srt/configs/nano_nemotron_vl.py

@@ -150,3 +150,12 @@ class NemotronH_Nano_VL_V2_Config(PretrainedConfig):
             video_maintain_aspect_ratio=self.video_maintain_aspect_ratio,
         )
         return radio_config
+
+
+class NemotronH_Nano_Omni_Reasoning_V3_Config(NemotronH_Nano_VL_V2_Config):
+    model_type = "NemotronH_Nano_Omni_Reasoning_V3"
+
+    def __init__(self, *args, **kwargs):
+        # Explicit __init__ prevents PretrainedConfig.__init_subclass__ from
+        # replacing the parent's custom __init__ with a dataclass-generated one.
+        super().__init__(*args, **kwargs)

python/sglang/srt/models/nano_nemotron_vl.py

@@ -372,4 +372,8 @@ class NemotronH_Nano_VL_V2(EVS):
             self.sound_encoder.load_weights(sound_weights)
 
 
-EntryClass = [NemotronH_Nano_VL_V2]
+class NemotronH_Nano_Omni_Reasoning_V3(NemotronH_Nano_VL_V2):
+    pass
+
+
+EntryClass = [NemotronH_Nano_VL_V2, NemotronH_Nano_Omni_Reasoning_V3]

python/sglang/srt/multimodal/processors/nano_nemotron_vl.py

@@ -19,13 +19,19 @@ import numpy as np
 import torch
 from PIL import Image
 
-from sglang.srt.configs.nano_nemotron_vl import NemotronH_Nano_VL_V2_Config
+from sglang.srt.configs.nano_nemotron_vl import (
+    NemotronH_Nano_Omni_Reasoning_V3_Config,
+    NemotronH_Nano_VL_V2_Config,
+)
 from sglang.srt.managers.schedule_batch import (
     Modality,
     MultimodalDataItem,
     MultimodalProcessorOutput,
 )
-from sglang.srt.models.nano_nemotron_vl import NemotronH_Nano_VL_V2
+from sglang.srt.models.nano_nemotron_vl import (
+    NemotronH_Nano_Omni_Reasoning_V3,
+    NemotronH_Nano_VL_V2,
+)
 from sglang.srt.models.parakeet import ParakeetExtractor
 from sglang.srt.multimodal.audio_from_video import extract_audio_from_video_bytes
 from sglang.srt.multimodal.evs import EVSProcessor
@@ -51,7 +57,7 @@ MAX_FRAMES = 128
 
 
 class NanoNemotronVLImageProcessor(BaseMultimodalProcessor):
-    models = [NemotronH_Nano_VL_V2]
+    models = [NemotronH_Nano_VL_V2, NemotronH_Nano_Omni_Reasoning_V3]
     gpu_image_decode = (
         False  # NanoNemotronVL processes loaded image as PIL image explicitly
     )
@@ -59,7 +65,11 @@ class NanoNemotronVLImageProcessor(BaseMultimodalProcessor):
     def __init__(self, hf_config, server_args, _image_processor, *args, **kwargs):
         super().__init__(hf_config, server_args, _image_processor, *args, **kwargs)
         self.evs = EVSProcessor(
-            hf_config, {NemotronH_Nano_VL_V2_Config: NemotronH_Nano_VL_V2}
+            hf_config,
+            {
+                NemotronH_Nano_VL_V2_Config: NemotronH_Nano_VL_V2,
+                NemotronH_Nano_Omni_Reasoning_V3_Config: NemotronH_Nano_Omni_Reasoning_V3,
+            },
         )
         Image.MAX_IMAGE_PIXELS = None
         self.image_size = hf_config.image_size

python/sglang/srt/utils/hf_transformers/common.py

@@ -39,6 +39,7 @@ from sglang.srt.configs import (
     KimiVLConfig,
     LongcatFlashConfig,
     MultiModalityConfig,
+    NemotronH_Nano_Omni_Reasoning_V3_Config,
     NemotronH_Nano_VL_V2_Config,
     NemotronHConfig,
     Olmo3Config,
@@ -87,6 +88,7 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
         DotsVLMConfig,
         DotsOCRConfig,
         NemotronH_Nano_VL_V2_Config,
+        NemotronH_Nano_Omni_Reasoning_V3_Config,
         NemotronHConfig,
         DeepseekVLV2Config,
         Qwen3_5Config,