fix: fix SHM pointer re-serialization in DP attention (#17930)

python/sglang/srt/managers/mm_utils.py

@@ -1513,6 +1513,27 @@ class ShmPointerMMData:
             self.shm.close()
 
     def __getstate__(self):
+        if not hasattr(self, "shm") or self.shm is None:
+            tensor = getattr(self, "cpu_tensor", None)
+            if tensor is None:
+                tensor = getattr(self, "tensor", None)
+            if tensor is None:
+                raise RuntimeError(
+                    "ShmPointerMMData cannot recreate shared memory without tensor"
+                )
+
+            cpu_tensor = tensor.cpu().contiguous()
+            self.shape = cpu_tensor.shape
+            self.dtype = cpu_tensor.dtype
+
+            nbytes = cpu_tensor.numel() * cpu_tensor.element_size()
+            self.shm = shared_memory.SharedMemory(create=True, size=nbytes)
+            try:
+                shm_view = np.ndarray((nbytes,), dtype=np.uint8, buffer=self.shm.buf)
+                shm_view[:] = cpu_tensor.view(torch.uint8).numpy().flatten()
+            finally:
+                self.shm.close()
+
         return {
             "shm_name": self.shm.name,
             "shape": self.shape,
@@ -1521,12 +1542,15 @@ class ShmPointerMMData:
 
     def __setstate__(self, state):
         self.shm_name = state["shm_name"]
+        self.shape = state["shape"]
+        self.dtype = state["dtype"]
+        self.shm = None
 
         shm_handle = shared_memory.SharedMemory(name=self.shm_name)
         try:
             self.tensor = (
-                torch.frombuffer(shm_handle.buf, dtype=state["dtype"])
-                .reshape(state["shape"])
+                torch.frombuffer(shm_handle.buf, dtype=self.dtype)
+                .reshape(self.shape)
                 .clone()
             )
         finally:

test/registered/distributed/test_dp_attention.py

@@ -3,6 +3,7 @@ from types import SimpleNamespace
 
 import requests
 
+from sglang.lang.chat_template import get_chat_template_by_model_path
 from sglang.srt.environ import envs
 from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import register_cuda_ci
@@ -13,6 +14,7 @@ from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
 from sglang.test.kits.regex_constrained_kit import TestRegexConstrainedMixin
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
+    DEFAULT_IMAGE_URL,
     DEFAULT_MLA_MODEL_NAME_FOR_TEST,
     DEFAULT_MODEL_NAME_FOR_TEST_MLA,
     DEFAULT_MODEL_NAME_FOR_TEST_MLA_NEXTN,
@@ -182,5 +184,52 @@ class TestDPAttentionDP2TP2DeepseekV3MTP(
         self.assertGreater(avg_spec_accept_length, 2.5)
 
 
+class TestDPAttentionDP2TP2VLM(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        # TODO(FlamingoPg): Use Kimi-VL-A3B-Instruct temporarily
+        # cauz Qwen3-VL use mrope which has bug in DP attention mode
+        cls.model = "moonshotai/Kimi-VL-A3B-Instruct"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.image_url = DEFAULT_IMAGE_URL
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--trust-remote-code",
+                "--tp",
+                "2",
+                "--enable-dp-attention",
+                "--dp",
+                "2",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_vlm_generate(self):
+        chat_template = get_chat_template_by_model_path(self.model)
+        prompt = f"{chat_template.image_token}What is in this image?"
+        response = requests.post(
+            self.base_url + "/generate",
+            json={
+                "text": prompt,
+                "image_data": [self.image_url],
+                "sampling_params": {
+                    "temperature": 0,
+                    "max_new_tokens": 16,
+                },
+            },
+        )
+        response.raise_for_status()
+        response_json = response.json()
+        print(response_json)
+        self.assertIn("output_ids", response_json)
+        self.assertGreater(len(response_json["output_ids"]), 0)
+
+
 if __name__ == "__main__":
     unittest.main()