[NemotronH] V3 Omni wrapper: WeightsMapper + config round-trip (#25023)

Co-authored-by: Yihao Wang <42559837+AgainstEntropy@users.noreply.github.com>

python/sglang/srt/configs/nano_nemotron_vl.py

@@ -60,6 +60,13 @@ class NemotronH_Nano_VL_V2_Config(PretrainedConfig):
         use_thumbnail: bool = True,
         **kwargs,
     ):
+        # Round-trip: `to_dict()` emits `raw_vision_config` (V2's storage
+        # name) but `from_dict()` rebuilds via this `vision_config` kwarg.
+        # Without this alias, the V3->V2 alias rebuild in `get_config` loses
+        # the vision config across the round-trip.
+        if vision_config is None:
+            vision_config = kwargs.pop("raw_vision_config", None)
+
         super().__init__(**kwargs)
 
         # Handle both cases: when loading from JSON (llm_config is dict) and when called internally by transformers (llm_config; vision_config are None)

python/sglang/srt/models/nano_nemotron_vl.py

@@ -39,6 +39,7 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.nemotron_h import NemotronHForCausalLM
 from sglang.srt.models.parakeet import ProjectedParakeet
 from sglang.srt.models.radio import RadioModel
+from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.multimodal.evs import EVS, EVSConfig
 from sglang.srt.multimodal.evs.evs_module import VideoEVSDataItem
 from sglang.srt.utils import add_prefix
@@ -47,6 +48,15 @@ logger = logging.getLogger(__name__)
 
 
 class NemotronH_Nano_VL_V2(EVS):
+    # The loader reads `hf_to_sglang_mapper` off the outer model class when
+    # applying name rewrites to the quant config's `quantized_layers` keys;
+    # the inner NemotronHForCausalLM mapper is not consulted there.
+    hf_to_sglang_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            "language_model.backbone.": "language_model.model.",
+        },
+    )
+
     @staticmethod
     def create_evs_config(config: NemotronH_Nano_VL_V2_Config):
         return EVSConfig(video_pruning_rate=config.video_pruning_rate)