[VLM] Revise InternVL Piecewise CUDA Graph Supporting (#13846)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2025-11-24 22:15:10 +08:00
parent ecefc7904f
commit 8ef11569a2
1 changed files with 4 additions and 0 deletions
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -362,6 +362,10 @@ class ModelRunner:
                    elif hasattr(layer.self_attn, "attn_mqa"):
                        # For DeepSeek model
                        self.attention_layers.append(layer.self_attn.attn_mqa)
+                # For InternVL model
+                elif hasattr(layer, "attention"):
+                    if hasattr(layer.attention, "attn"):
+                        self.attention_layers.append(layer.attention.attn)

            if len(self.attention_layers) < self.model_config.num_hidden_layers:
                # TODO(yuwei): support Non-Standard GQA