Bug fix for llama eagle3 (#21397)

python/sglang/srt/models/llama.py

@@ -252,8 +252,13 @@ class LlamaDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
-        rope_scaling = config.rope_parameters
+        rope_parameters = getattr(config, "rope_parameters", None)
+        if rope_parameters is not None:
+            rope_theta = rope_parameters.get("rope_theta", 10000)
+            rope_scaling = rope_parameters
+        else:
+            rope_theta = getattr(config, "rope_theta", 10000)
+            rope_scaling = getattr(config, "rope_scaling", None)
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None
         ):

python/sglang/srt/models/llama_eagle3.py

@@ -111,13 +111,17 @@ class LlamaModel(nn.Module):
         super().__init__()
         self.config = config
 
-        rope_scaling = config.rope_parameters
+        rope_parameters = getattr(config, "rope_parameters", None)
+        if rope_parameters is not None:
+            rope_scaling = rope_parameters
+        else:
+            rope_scaling = getattr(config, "rope_scaling", None)
         self.is_mrope_enabled = (
             rope_scaling is not None and "mrope_section" in rope_scaling
         )
         # fix rope_scaling for qwen2.5-vl
         if self.is_mrope_enabled:
-            config.rope_parameters["rope_type"] = "default"
+            rope_scaling["rope_type"] = "default"
 
         self.vocab_size = config.vocab_size
         self.embed_tokens = VocabParallelEmbedding(