Support piecewise cuda graph for Qwen3-next (#13081)

2025-11-25 21:01:27 +08:00
parent 432ecf841e
commit d64bf6c6ce
6 changed files with 112 additions and 3 deletions
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -363,6 +363,11 @@ class ModelRunner:
                    elif hasattr(layer.self_attn, "attn_mqa"):
                        # For DeepSeek model
                        self.attention_layers.append(layer.self_attn.attn_mqa)
+                # For hybrid model
+                elif hasattr(layer, "attn"):
+                    self.attention_layers.append(layer.attn)
+                elif hasattr(layer, "linear_attn"):
+                    self.attention_layers.append(layer.linear_attn)
                # For InternVL model
                elif hasattr(layer, "attention"):
                    if hasattr(layer.attention, "attn"):