Support GlmMoeDsaForCausalLM (#18521)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Signed-off-by: BBuf <1182563586@qq.com>
Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
Co-authored-by: BBuf <1182563586@qq.com>

python/sglang/srt/configs/model_config.py

@@ -61,6 +61,7 @@ def is_deepseek_nsa(config: PretrainedConfig) -> bool:
             "DeepseekV3ForCausalLMNextN",
             "MistralLarge3ForCausalLM",
             "PixtralForConditionalGeneration",
+            "GlmMoeDsaForCausalLM",
         ]
         and getattr(config, "index_topk", None) is not None
     )
@@ -271,10 +272,10 @@ class ModelConfig:
     def _config_draft_model(self):
         is_draft_model = self.is_draft_model
 
-        if (
-            is_draft_model
-            and self.hf_config.architectures[0] == "DeepseekV3ForCausalLM"
-        ):
+        if is_draft_model and self.hf_config.architectures[0] in [
+            "DeepseekV3ForCausalLM",
+            "GlmMoeDsaForCausalLM",
+        ]:
             self.hf_config.architectures[0] = "DeepseekV3ForCausalLMNextN"
 
         if is_draft_model and self.hf_config.architectures[0] in [
@@ -411,7 +412,6 @@ class ModelConfig:
             "swa_v_head_dim",
             self.v_head_dim,
         )
-
         # FIXME: temporary special judge for MLA architecture
         if (
             "DeepseekV2ForCausalLM" in self.hf_config.architectures
@@ -419,6 +419,7 @@ class ModelConfig:
             or "DeepseekV3ForCausalLM" in self.hf_config.architectures
             or "DeepseekV3ForCausalLMNextN" in self.hf_config.architectures
             or "Glm4MoeLiteForCausalLM" in self.hf_config.architectures
+            or "GlmMoeDsaForCausalLM" in self.hf_config.architectures
             or "LongcatFlashForCausalLM" in self.hf_config.architectures
             or "LongcatFlashForCausalLMNextN" in self.hf_config.architectures
             or "DotsVLMForCausalLM" in self.hf_config.architectures