Fix CUDA 13.0 VMM handle type compatibility (#34431)

python/sglang/srt/cuda_vmm_utils.py

@@ -281,20 +281,28 @@ def make_device_allocation_prop(
     elif not isinstance(handle_types, int):
         raise ValueError("handle_types must be 'auto', an integer, or None")
 
-    handle_types = int(handle_types)
+    handle_type_value = int(handle_types)
     valid_handle_types = {
-        int(drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_NONE),
+        int(drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_NONE): (
-        int(drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR),
+            drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_NONE
-        int(drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_FABRIC),
+        ),
+        int(drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR): (
+            drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR
+        ),
+        int(drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_FABRIC): (
+            drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_FABRIC
+        ),
     }
-    if handle_types not in valid_handle_types:
+    if handle_type_value not in valid_handle_types:
-        raise ValueError(f"invalid CUDA handle-type value: {handle_types}")
+        raise ValueError(f"invalid CUDA handle-type value: {handle_type_value}")
 
     prop = drv.CUmemAllocationProp()
     prop.type = drv.CUmemAllocationType.CU_MEM_ALLOCATION_TYPE_PINNED
     prop.location.type = drv.CUmemLocationType.CU_MEM_LOCATION_TYPE_DEVICE
     prop.location.id = int(device_id)
-    prop.requestedHandleTypes = handle_types
+    # cuda-bindings 13.0.x requires the generated enum here; newer releases
+    # also accept a plain int, which previously hid this compatibility issue.
+    prop.requestedHandleTypes = valid_handle_types[handle_type_value]
     prop.allocFlags.gpuDirectRDMACapable = int(gpu_direct_rdma)
     return prop
 

test/registered/unit/test_cuda_vmm_utils.py

@@ -159,6 +159,46 @@ def test_default_handle_type_fallback(monkeypatch, rejected, expected) -> None:
         get_device_allocation_handle_type.cache_clear()
 
 
+@pytest.mark.parametrize(
+    ("handle_types", "expected"),
+    [
+        (None, drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_NONE),
+        (0, drv.CUmemAllocationHandleType.CU_MEM_HANDLE_TYPE_NONE),
+        (_POSIX_FD, _POSIX_FD),
+        (_FABRIC, _FABRIC),
+    ],
+)
+def test_allocation_prop_assigns_handle_type_enum(
+    monkeypatch, handle_types, expected
+) -> None:
+    """CUDA bindings 13.0.x reject integer VMM handle types at assignment."""
+
+    class Fields:
+        pass
+
+    class StrictAllocationProp:
+        def __init__(self):
+            self.location = Fields()
+            self.allocFlags = Fields()
+            self._requested_handle_types = None
+
+        @property
+        def requestedHandleTypes(self):
+            return self._requested_handle_types
+
+        @requestedHandleTypes.setter
+        def requestedHandleTypes(self, value):
+            if not isinstance(value, drv.CUmemAllocationHandleType):
+                raise TypeError("requestedHandleTypes requires a CUDA enum")
+            self._requested_handle_types = value
+
+    monkeypatch.setattr(drv, "CUmemAllocationProp", StrictAllocationProp)
+
+    prop = make_device_allocation_prop(0, handle_types=handle_types)
+
+    assert prop.requestedHandleTypes is expected
+
+
 def test_granularity_defaults_to_recommended(monkeypatch) -> None:
     prop = make_device_allocation_prop(0, handle_types=None)
     seen = []