refactor: multimodal data (#4754)

python/sglang/srt/models/deepseek_v2.py

@@ -1308,6 +1308,9 @@ class DeepseekV2ForCausalLM(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.dp_size = get_attention_dp_size()
 
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.model.embed_tokens
+
     @torch.no_grad()
     def forward(
         self,