[VLM] Support Piecewise CUDA Graph for Qwen3-Omni-MOE (#14222)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>

python/sglang/srt/model_executor/model_runner.py

@@ -28,6 +28,7 @@ from typing import Callable, List, Optional, Tuple, Union
 
 import torch
 import torch.distributed as dist
+from torch import nn
 
 from sglang.srt.configs import (
     FalconH1Config,
@@ -247,6 +248,13 @@ if _is_npu:
     torch_npu.npu.set_compile_mode(jit_compile=False)
 
 
+def resolve_language_model(model: nn.Module) -> nn.Module:
+    model_cls_name = model.__class__.__name__
+    if model_cls_name == "Qwen3OmniMoeForConditionalGeneration":
+        return model.thinker.model
+    return model.model
+
+
 class RankZeroFilter(logging.Filter):
     """Filter that only allows INFO level logs from rank 0, but allows all other levels from any rank."""
 
@@ -2434,6 +2442,7 @@ class ModelRunner:
 
         # Collect attention layers from the model
         self.attention_layers = []
+        self.model.model = resolve_language_model(self.model)
         for layer in self.model.model.layers:
             if hasattr(layer, "self_attn"):
                 if hasattr(layer.self_attn, "attn"):

test/srt/run_suite.py

@@ -97,7 +97,6 @@ suites = {
         TestFile("test_original_logprobs.py", 41),
         TestFile("test_page_size.py", 60),
         TestFile("test_penalty.py", 82),
-        TestFile("test_piecewise_cuda_graph.py", 850),
         TestFile("test_priority_scheduling.py", 130),
         TestFile("test_pytorch_sampling_backend.py", 66),
         TestFile("test_radix_attention.py", 105),
@@ -158,6 +157,7 @@ suites = {
         TestFile("test_local_attn.py", 411),
         TestFile("test_multi_instance_release_memory_occupation.py", 64),
         TestFile("test_pp_single_node.py", 481),
+        TestFile("test_piecewise_cuda_graph.py", 1200),
     ],
     "per-commit-8-gpu-h200": [
         TestFile("test_deepseek_v3_basic.py", 275),

test/srt/test_piecewise_cuda_graph.py

@@ -371,5 +371,47 @@ class TestPiecewiseCudaGraphQwen25VLEmbedding(CustomTestCase):
         )
 
 
+class TestPiecewiseCudaGraphQwen3OmniMOE(CustomTestCase):
+    """Test piecewise CUDA graph with Qwen3-Omni-30B-A3B-Instruct  model"""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--enable-piecewise-cuda-graph",
+                "--piecewise-cuda-graph-compiler",
+                "eager",
+                "--disable-radix-cache",
+                "--tp=4",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k_accuracy(self):
+        """Test GSM8K accuracy with 8-shot setting"""
+        num_examples = 2000
+
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mgsm_en",
+            num_examples=num_examples,
+            num_threads=min(num_examples, 1024),
+        )
+
+        metrics = run_eval(args)
+        print(f"GSM8K Accuracy: {metrics['score']:.3f}")
+
+        self.assertGreaterEqual(metrics["score"], 0.70)
+
+
 if __name__ == "__main__":
     unittest.main()