model: Support Hybrid Mamba2 NemotronHForCausalLM (nvidia/NVIDIA-Nemotron-Nano-9B-v2) (#10909)

Signed-off-by: Netanel Haber <nhaber@nvidia.com>

python/sglang/srt/utils/common.py

@@ -518,6 +518,24 @@ def make_layers(
     return modules, start_layer, end_layer
 
 
+def make_layers_non_pp(
+    num_hidden_layers: int,
+    layer_fn: LayerFn,
+    prefix: str = "",
+) -> torch.nn.ModuleList:
+    from sglang.srt.offloader import get_offloader
+
+    layers = torch.nn.ModuleList(
+        get_offloader().wrap_modules(
+            (
+                layer_fn(idx=idx, prefix=add_prefix(idx, prefix))
+                for idx in range(num_hidden_layers)
+            )
+        )
+    )
+    return layers
+
+
 cmo_stream = None
 
 

python/sglang/srt/utils/hf_transformers_utils.py

@@ -45,6 +45,7 @@ from sglang.srt.configs import (
     KimiVLConfig,
     LongcatFlashConfig,
     MultiModalityConfig,
+    NemotronHConfig,
     Qwen3NextConfig,
     Step3VLConfig,
 )
@@ -66,6 +67,7 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
     FalconH1Config.model_type: FalconH1Config,
     DotsVLMConfig.model_type: DotsVLMConfig,
     DotsOCRConfig.model_type: DotsOCRConfig,
+    NemotronHConfig.model_type: NemotronHConfig,
 }
 
 for name, cls in _CONFIG_REGISTRY.items():