diff --git a/csrc/cumem_allocator.cpp b/csrc/cumem_allocator.cpp
index 720c990450..6965d2124a 100644
--- a/csrc/cumem_allocator.cpp
+++ b/csrc/cumem_allocator.cpp
@@ -203,12 +203,25 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
   }
 #endif
 
-  CUmemAccessDesc accessDesc = {};
-  accessDesc.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-  accessDesc.location.id = device;
-  accessDesc.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+#if defined(USE_ROCM) && HIP_VERSION >= 70200000
+  // Also map for the host, as PyTorch does for its expandable segments: on
+  // ROCm >= 7.2, Tensor.item() on a large-BAR device dereferences VRAM
+  // directly from the host, which segfaults on a device-only mapping.
+  constexpr int num_desc = 2;
+#else
+  constexpr int num_desc = 1;
+#endif
+  CUmemAccessDesc accessDesc[num_desc] = {};
+  accessDesc[0].location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+  accessDesc[0].location.id = device;
+  accessDesc[0].flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+#if defined(USE_ROCM) && HIP_VERSION >= 70200000
+  accessDesc[1].location.type = hipMemLocationTypeHost;
+  accessDesc[1].location.id = 0;  // ignored
+  accessDesc[1].flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+#endif
 
-  CUDA_CHECK(cuMemSetAccess(d_mem, size, &accessDesc, 1));
+  CUDA_CHECK(cuMemSetAccess(d_mem, size, accessDesc, num_desc));
   if (error_code != 0) {
     return;
   }
diff --git a/tests/basic_correctness/memory/sleep_mode/test_sleep_mode.py b/tests/basic_correctness/memory/sleep_mode/test_sleep_mode.py
index 73af6e3910..9f6180335a 100644
--- a/tests/basic_correctness/memory/sleep_mode/test_sleep_mode.py
+++ b/tests/basic_correctness/memory/sleep_mode/test_sleep_mode.py
@@ -323,7 +323,6 @@ def _attn_kv_scales(model) -> dict[str, list[float]]:
     }
 
 
-@pytest.mark.skipif(not current_platform.is_cuda(), reason="Segfaults on ROCm.")
 @requires_fp8
 @create_new_process_for_each_test()
 def test_deep_sleep_compressed_tensors_kv_scales(