Fix EAGLE3 for llama3.3 70b (#4716)

python/sglang/srt/models/llama.py

@@ -610,6 +610,12 @@ class LlamaForCausalLM(nn.Module):
         return self.model.embed_tokens.weight
 
     def set_embed(self, embed):
+        # NOTE: If draft hidden size != target hidden size, the embed weight cannot be shared for EAGLE3
+        if (
+            hasattr(self.config, "target_hidden_size")
+            and self.config.target_hidden_size != self.config.hidden_size
+        ):
+            return
         del self.model.embed_tokens.weight
         self.model.embed_tokens.weight = embed
         torch.cuda.empty_cache()

python/sglang/srt/models/llama_eagle3.py

@@ -105,6 +105,9 @@ class LlamaModel(nn.Module):
             prefix=add_prefix("embed_tokens", prefix),
         )
         self.midlayer = LlamaDecoderLayer(config, 0, quant_config, prefix)
+        if hasattr(config, "target_hidden_size"):
+            self.fc = torch.nn.Linear(config.target_hidden_size * 3, config.hidden_size)
+        else:
             self.fc = torch.nn.Linear(config.hidden_size * 3, config.hidden_size)
 
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)