Add registration API for external linear attention backend (#21983)

Signed-off-by: Xingyu Liu <charlotteliu12x@gmail.com>

python/sglang/srt/server_args.py

@@ -26,6 +26,7 @@ import random
 import tempfile
 from typing import Any, Callable, Dict, List, Literal, Optional, Union
 
+from sglang.srt.configs.linear_attn_model_registry import get_linear_attn_spec_by_arch
 from sglang.srt.connector import ConnectorType
 from sglang.srt.environ import envs
 from sglang.srt.function_call.function_call_parser import FunctionCallParser
@@ -1501,6 +1502,14 @@ class ServerArgs:
         hf_config = self.get_model_config().hf_config
         model_arch = hf_config.architectures[0]
 
+        _hybrid_spec = get_linear_attn_spec_by_arch(model_arch)
+        if _hybrid_spec is not None:
+            self._handle_mamba_radix_cache(
+                model_arch=model_arch,
+                support_mamba_cache=_hybrid_spec.support_mamba_cache,
+                support_mamba_cache_extra_buffer=_hybrid_spec.support_mamba_cache_extra_buffer,
+            )
+
         if model_arch in [
             "MistralLarge3ForCausalLM",
             "PixtralForConditionalGeneration",