diff --git a/src/cco/cco_init.cpp b/src/cco/cco_init.cpp
index b8f1f57ab..622cc685b 100644
--- a/src/cco/cco_init.cpp
+++ b/src/cco/cco_init.cpp
@@ -99,7 +99,29 @@ static hipMemAllocationType CcoWindowAllocType() {
     const char* e = getenv("CCO_UNCACHED_WINDOW");
     return e && atoi(e) == 0;
   }();
-  return cached ? hipMemAllocationTypePinned : hipMemAllocationTypeUncached;
+  if (cached) return hipMemAllocationTypePinned;
+  // hipMemAllocationTypeUncached was added on 2025-09-11 (HIP_VERSION > 70051831).
+  // Guard the symbol so CCO still builds on older ROCm (e.g. 7.0.0); mirror the
+  // shmem path in symmetric_memory.cpp: use the raw value 0x40000000 on the
+  // 70051831 build where the symbol may be missing (excluding the buggy
+  // 7c9236b16 build), and fall back to Pinned on anything older.
+#if HIP_VERSION > 70051831
+  return hipMemAllocationTypeUncached;
+#elif HIP_VERSION == 70051831
+  if (strcmp(HIP_VERSION_GITHASH, "7c9236b16") != 0) {
+    return static_cast<hipMemAllocationType>(0x40000000);  // hipMemAllocationTypeUncached
+  }
+  MORI_SHMEM_WARN(
+      "CCO uncached window requested but ROCm build {} has a known "
+      "hipMemAllocationTypeUncached issue; falling back to Pinned memory",
+      HIP_VERSION_GITHASH);
+  return hipMemAllocationTypePinned;
+#else
+  MORI_SHMEM_WARN(
+      "CCO uncached window requested but ROCm version does not support "
+      "hipMemAllocationTypeUncached; falling back to Pinned memory");
+  return hipMemAllocationTypePinned;
+#endif
 }
 
 // Local slot base = the VA where this rank's slice of the flat VA starts.
diff --git a/tests/python/ops/dispatch_combine_v2/bench_dispatch_combine.py b/tests/python/ops/dispatch_combine_v2/bench_dispatch_combine.py
index edd28c88c..8653930e8 100644
--- a/tests/python/ops/dispatch_combine_v2/bench_dispatch_combine.py
+++ b/tests/python/ops/dispatch_combine_v2/bench_dispatch_combine.py
@@ -279,6 +279,11 @@ def mock_fmoe():
         def time_eager(fn, ct):
             for _ in range(WARMUP):
                 fn(ct)
+                sync()
+                comm.barrier()  # lock-step warmup: sync(device)+barrier so no
+                # rank starts call N+1 before all finish N (avoids the cco cross-
+                # device barrier flag-overwrite deadlock at small token counts).
+                # Untimed, so it does not affect the measured replay bandwidth.
             sync()
             comm.barrier()
             s, e = torch.cuda.Event(enable_timing=True), torch.cuda.Event(
@@ -295,6 +300,11 @@ def time_eager(fn, ct):
         def time_graph(fn, ct):
             for _ in range(WARMUP):
                 fn(ct)
+                sync()
+                comm.barrier()  # lock-step warmup: sync(device)+barrier so no
+                # rank starts call N+1 before all finish N (avoids the cco cross-
+                # device barrier flag-overwrite deadlock at small token counts).
+                # Untimed, so it does not affect the measured replay bandwidth.
             sync()
             gr = torch.cuda.CUDAGraph()
             with torch.cuda.graph(gr):