[Bug][VLM] Fix shared memory race condition in ShmPointerMMData broadcast for multi-GPU VLM serving (#21655)

python/sglang/srt/managers/mm_utils.py

@@ -1725,6 +1725,19 @@ def wrap_shm_features(obj):
     return obj
 
 
+def has_shm_features(recv_reqs):
+    """Return True if any request in the list contains ShmPointerMMData."""
+    for req in recv_reqs:
+        if hasattr(req, "batch"):
+            if has_shm_features(req.batch):
+                return True
+        elif hasattr(req, "mm_inputs") and req.mm_inputs:
+            for item in req.mm_inputs.get("mm_items", []):
+                if isinstance(item.feature, ShmPointerMMData):
+                    return True
+    return False
+
+
 def unwrap_shm_features(obj):
     """
     Restore ShmPointerMMData wrappers back into standard torch.Tensors.

python/sglang/srt/managers/scheduler.py

@@ -142,7 +142,11 @@ from sglang.srt.managers.io_struct import (
     UpdateWeightsFromIPCReqInput,
     UpdateWeightsFromTensorReqInput,
 )
-from sglang.srt.managers.mm_utils import init_mm_embedding_cache, unwrap_shm_features
+from sglang.srt.managers.mm_utils import (
+    has_shm_features,
+    init_mm_embedding_cache,
+    unwrap_shm_features,
+)
 from sglang.srt.managers.multimodal_processor import get_mm_processor, import_processors
 from sglang.srt.managers.overlap_utils import FutureMap
 from sglang.srt.managers.prefill_delayer import (
@@ -1501,6 +1505,27 @@ class Scheduler(
         # so that ShmPointerMMData metadata (not full tensor data) is what
         # gets serialized during broadcast_pyobj.
         if recv_reqs:
+            # Barrier for the non-DP-attention path only: there is a single
+            # broadcast_pyobj on tp_cpu_group where the source rank returns
+            # the original objects immediately while other ranks are still in
+            # pickle.loads (-> __setstate__ -> shm_open).  Without a barrier
+            # the source can call materialize() / shm_unlink before others
+            # open the segment.  recv_reqs is consistent across all ranks
+            # here (same broadcast), so the guard is deadlock-free.
+            #
+            # Under DP-attention no barrier is needed: the control_reqs
+            # broadcast on tp_cpu_group (step 3) is a collective that forces
+            # every rank to complete the earlier attn_tp / attn_cp work_reqs
+            # deserializations (steps 1-2, which call shm_open) before any
+            # rank returns from step 3.  POSIX guarantees shm_unlink only
+            # removes the name; already-open handles stay valid.
+            if (
+                not self.server_args.enable_dp_attention
+                and self.tp_size > 1
+                and self.model_config.is_multimodal
+                and has_shm_features(recv_reqs)
+            ):
+                barrier(group=self.tp_cpu_group)
             for req in recv_reqs:
                 unwrap_shm_features(req)