introduce arg_groups/ with nemotron_h hook (#24328)

python/sglang/srt/arg_groups/nemotron_h_hook.py

@@ -0,0 +1,51 @@
+import logging
+from typing import TYPE_CHECKING
+
+from sglang.srt.utils.common import is_sm100_supported
+
+if TYPE_CHECKING:
+    from sglang.srt.server_args import ServerArgs
+
+logger = logging.getLogger(__name__)
+
+
+def apply_nemotron_h_defaults(server_args: "ServerArgs", model_arch: str) -> None:
+    """Apply NemotronH model-specific server arg defaults and constraints."""
+    model_config = server_args.get_model_config()
+    if model_config.quantization in [
+        "modelopt",
+        "modelopt_fp8",
+        "modelopt_fp4",
+        "modelopt_mixed",
+    ]:
+        assert model_config.hf_config.mlp_hidden_act == "relu2"
+        if model_config.quantization == "modelopt":
+            quant_algo = model_config.hf_config.quantization_config["quant_algo"]
+            if quant_algo == "MIXED_PRECISION":
+                server_args.quantization = "modelopt_mixed"
+            else:
+                server_args.quantization = (
+                    "modelopt_fp4" if quant_algo == "NVFP4" else "modelopt_fp8"
+                )
+        else:
+            server_args.quantization = model_config.quantization
+        if server_args.moe_runner_backend == "auto":
+            if is_sm100_supported() and server_args.moe_a2a_backend == "none":
+                server_args.moe_runner_backend = "flashinfer_trtllm"
+                logger.info(
+                    "Use flashinfer_trtllm as MoE runner backend on sm100 for "
+                    f"{model_arch}"
+                )
+            else:
+                server_args.moe_runner_backend = "flashinfer_cutlass"
+
+    server_args._handle_mamba_radix_cache(
+        model_arch=model_arch,
+        support_mamba_cache=True,
+        support_mamba_cache_extra_buffer=False,
+        sm100_default_attention_backend="flashinfer",
+    )
+    assert server_args.attention_backend != "triton", (
+        "NemotronHForCausalLM does not support triton attention backend,"
+        "as the first layer might not be an attention layer"
+    )

python/sglang/srt/server_args.py

@@ -2143,46 +2143,11 @@ class ServerArgs:
                 support_mamba_cache=False,
             )
         elif model_arch in ["NemotronHForCausalLM"]:
-            model_config = self.get_model_config()
+            from sglang.srt.arg_groups.nemotron_h_hook import (
-            if model_config.quantization in [
+                apply_nemotron_h_defaults,
-                "modelopt",
-                "modelopt_fp8",
-                "modelopt_fp4",
-                "modelopt_mixed",
-            ]:
-                assert model_config.hf_config.mlp_hidden_act == "relu2"
-                if model_config.quantization == "modelopt":
-                    quant_algo = model_config.hf_config.quantization_config[
-                        "quant_algo"
-                    ]
-                    if quant_algo == "MIXED_PRECISION":
-                        self.quantization = "modelopt_mixed"
-                    else:
-                        self.quantization = (
-                            "modelopt_fp4" if quant_algo == "NVFP4" else "modelopt_fp8"
             )
-                else:
-                    self.quantization = model_config.quantization
-                if self.moe_runner_backend == "auto":
-                    if is_sm100_supported() and self.moe_a2a_backend == "none":
-                        self.moe_runner_backend = "flashinfer_trtllm"
-                        logger.info(
-                            "Use flashinfer_trtllm as MoE runner backend on sm100 for "
-                            f"{model_arch}"
-                        )
-                    else:
-                        self.moe_runner_backend = "flashinfer_cutlass"
 
-            self._handle_mamba_radix_cache(
+            apply_nemotron_h_defaults(self, model_arch)
-                model_arch=model_arch,
-                support_mamba_cache=True,
-                support_mamba_cache_extra_buffer=False,
-                sm100_default_attention_backend="flashinfer",
-            )
-            assert self.attention_backend != "triton", (
-                "NemotronHForCausalLM does not support triton attention backend,"
-                "as the first layer might not be an attention layer"
-            )
         elif model_arch in [
             "Qwen3MoeForCausalLM",
             "Qwen3VLMoeForConditionalGeneration",