[LFM2] Wire Lfm2MoeForCausalLM into the LFM2 serving override tables (#30780)

python/sglang/srt/arg_groups/overrides.py

@@ -774,7 +774,7 @@ def _granite_moe_hybrid_overrides(server_args: Any, hf_config: Any) -> dict:
     return {}
 
 
-@_register_for("Lfm2ForCausalLM")
+@_register_for("Lfm2ForCausalLM", "Lfm2MoeForCausalLM")
 def _lfm2_overrides(server_args: Any, hf_config: Any) -> dict:
     if is_sm100_supported() and server_args.attention_backend is None:
         return {"attention_backend": "flashinfer"}
@@ -1123,6 +1123,7 @@ _MAMBA_RADIX_CACHE_ARCHS = frozenset(
         "JetNemotronForCausalLM",
         "JetVLMForConditionalGeneration",
         "Lfm2ForCausalLM",
+        "Lfm2MoeForCausalLM",
         "ZayaForCausalLM",
     }
 )

python/sglang/srt/server_args.py

@@ -5395,7 +5395,7 @@ class ServerArgs:
             # _glm4_moe_overrides).
             pass
 
-        elif model_arch in ["Lfm2ForCausalLM"]:
+        elif model_arch in ["Lfm2ForCausalLM", "Lfm2MoeForCausalLM"]:
             # Attention backend selection moved to the override registry
             # (arg_groups/overrides.py: _lfm2_overrides).
             assert resolved_view(self).attention_backend != "triton", (