Preserve memory pool CUDA errors and harden OOM tests

cuda_bindings/tests/test_interoperability.py

@@ -6,6 +6,7 @@
 
 import cuda.bindings.driver as cuda
 import cuda.bindings.runtime as cudart
+from cuda.bindings._test_helpers.mempool import xfail_if_mempool_oom
 
 
 def supportsMemoryPool():
@@ -87,12 +88,14 @@ def test_interop_graphNode():
 def test_interop_memPool():
     # DRV to RT
     err_dr, pool = cuda.cuDeviceGetDefaultMemPool(0)
+    xfail_if_mempool_oom(err_dr, "cuDeviceGetDefaultMemPool", 0)
     assert err_dr == cuda.CUresult.CUDA_SUCCESS
     (err_rt,) = cudart.cudaDeviceSetMemPool(0, pool)
     assert err_rt == cudart.cudaError_t.cudaSuccess
 
     # RT to DRV
     err_rt, pool = cudart.cudaDeviceGetDefaultMemPool(0)
+    xfail_if_mempool_oom(err_rt, "cudaDeviceGetDefaultMemPool", 0)
     assert err_rt == cudart.cudaError_t.cudaSuccess
     (err_dr,) = cuda.cuDeviceSetMemPool(0, pool)
     assert err_dr == cuda.CUresult.CUDA_SUCCESS

cuda_core/cuda/core/_memory/_device_memory_resource.pyx

@@ -13,6 +13,7 @@ from cuda.core._memory._ipc cimport IPCAllocationHandle
 from cuda.core._resource_handles cimport (
     as_cu,
     get_device_mempool,
+    get_last_error,
 )
 from cuda.core._utils.cuda_utils cimport (
     check_or_create_options,
@@ -262,6 +263,14 @@ cdef inline _DMR_init(DeviceMemoryResource self, device_id, options):
 
     if opts is None:
         self._h_pool = get_device_mempool(dev_id)
+        if not self._h_pool:
+            HANDLE_RETURN(get_last_error())
+            raise RuntimeError(
+                f"Failed to initialize DeviceMemoryResource for device {dev_id}: "
+                "cuda-core returned an empty memory pool handle without recording a CUDA error. "
+                "This is an internal cuda-core error; please report it with your CUDA driver, "
+                "CUDA Toolkit, and cuda-python versions."
+            )
         self._mempool_owned = False
         MP_raise_release_threshold(self)
     else:

cuda_core/cuda/core/_memory/_graph_memory_resource.pyx

@@ -11,6 +11,7 @@ from cuda.core._memory._buffer cimport Buffer, Buffer_from_deviceptr_handle, Mem
 from cuda.core._resource_handles cimport (
     DevicePtrHandle,
     deviceptr_alloc_async,
+    get_last_error,
     as_cu,
 )
 
@@ -194,7 +195,13 @@ cdef inline Buffer GMR_allocate(cyGraphMemoryResource self, size_t size, Stream
         check_capturing(s)
         h_ptr = deviceptr_alloc_async(size, stream._h_stream)
     if not h_ptr:
-        raise RuntimeError("Failed to allocate memory asynchronously")
+        HANDLE_RETURN(get_last_error())
+        raise RuntimeError(
+            f"Failed to allocate {size} bytes from GraphMemoryResource: "
+            "cuda-core returned an empty allocation handle without recording a CUDA error. "
+            "This is an internal cuda-core error; please report it with your CUDA driver, "
+            "CUDA Toolkit, and cuda-python versions."
+        )
     return Buffer_from_deviceptr_handle(h_ptr, size, self, None)
 
 

cuda_core/cuda/core/_memory/_ipc.pyx

@@ -211,7 +211,13 @@ cdef _MemPool MP_from_allocation_handle(cls, alloc_handle):
     cdef int ipc_fd = int(alloc_handle)
     self._h_pool = create_mempool_handle_ipc(ipc_fd, IPC_HANDLE_TYPE)
     if not self._h_pool:
-        raise RuntimeError("Failed to import memory pool from IPC handle")
+        HANDLE_RETURN(get_last_error())
+        raise RuntimeError(
+            f"Failed to import {cls.__name__} from an allocation handle: "
+            "cuda-core returned an empty memory pool handle without recording a CUDA error. "
+            "This is an internal cuda-core error; please report it with your CUDA driver, "
+            "CUDA Toolkit, and cuda-python versions."
+        )
     self._ipc_data = IPCDataForMR(alloc_handle, True)
 
     # Register it.

cuda_core/cuda/core/_memory/_memory_pool.pyx

@@ -17,6 +17,7 @@ from cuda.core._resource_handles cimport (
     DevicePtrHandle,
     create_mempool_handle,
     deviceptr_alloc_from_pool,
+    get_last_error,
     as_cu,
     as_py,
 )
@@ -228,6 +229,14 @@ cdef int MP_init_create_pool(
 
     self._mempool_owned = True
     self._h_pool = create_mempool_handle(properties)
+    if not self._h_pool:
+        HANDLE_RETURN(get_last_error())
+        raise RuntimeError(
+            f"Failed to initialize {self.__class__.__name__}: "
+            "cuda-core returned an empty memory pool handle without recording a CUDA error. "
+            "This is an internal cuda-core error; please report it with your CUDA driver, "
+            "CUDA Toolkit, and cuda-python versions."
+        )
 
     if ipc_enabled:
         alloc_handle = _ipc.MP_export_mempool(self)
@@ -307,7 +316,13 @@ cdef inline Buffer _MP_allocate(_MemPool self, size_t size, Stream stream):
         check_not_capturing(s)
         h_ptr = deviceptr_alloc_from_pool(size, self._h_pool, stream._h_stream)
     if not h_ptr:
-        raise RuntimeError("Failed to allocate memory from pool")
+        HANDLE_RETURN(get_last_error())
+        raise RuntimeError(
+            f"Failed to allocate {size} bytes from {self.__class__.__name__}: "
+            "cuda-core returned an empty allocation handle without recording a CUDA error. "
+            "This is an internal cuda-core error; please report it with your CUDA driver, "
+            "CUDA Toolkit, and cuda-python versions."
+        )
     return Buffer_from_deviceptr_handle(h_ptr, size, self, None)
 
 

cuda_core/tests/conftest.py

@@ -5,6 +5,7 @@
 import os
 import pathlib
 import sys
+from contextlib import contextmanager
 from importlib.metadata import PackageNotFoundError, distribution
 
 import pytest
@@ -87,6 +88,8 @@ def create_managed_memory_resource_or_skip(*args, xfail_device=None, **kwargs):
         return ManagedMemoryResource(*args, **kwargs)
     except CUDAError as e:
         xfail_if_mempool_oom(e, _device_id_from_resource_options(xfail_device, args, kwargs))
+        if "CUDA_ERROR_NOT_SUPPORTED" in str(e):
+            pytest.skip("ManagedMemoryResource is not supported on this platform/device")
         raise
     except RuntimeError as e:
         if "requires CUDA 13.0" in str(e):
@@ -102,6 +105,15 @@ def create_pinned_memory_resource_or_xfail(*args, xfail_device=None, **kwargs):
         raise
 
 
+@contextmanager
+def xfail_on_graph_mempool_oom(device=0):
+    try:
+        yield
+    except CUDAError as e:
+        xfail_if_mempool_oom(e, "cuGraphAddMemAllocNode", device)
+        raise
+
+
 def _device_id_from_resource_options(device, args, kwargs):
     if device is not None:
         return device

cuda_core/tests/graph/test_graph_definition.py

@@ -10,6 +10,7 @@
 from helpers.graph_kernels import compile_common_kernels
 from helpers.misc import try_create_condition
 
+from conftest import xfail_on_graph_mempool_oom
 from cuda.core import Device, LaunchConfig
 from cuda.core.graph import (
     AllocNode,
@@ -201,13 +202,15 @@ def _build_disconnected():
 def graph_spec(request, init_cuda):
     if request.param is not _build_empty:
         _skip_if_no_mempool()
-    return request.param()
+    with xfail_on_graph_mempool_oom():
+        return request.param()
 
 
 @pytest.fixture(params=_NONEMPTY_BUILDERS)
 def nonempty_graph_spec(request, init_cuda):
     _skip_if_no_mempool()
-    return request.param()
+    with xfail_on_graph_mempool_oom():
+        return request.param()
 
 
 # =============================================================================
@@ -562,7 +565,8 @@ def node_spec(request, init_cuda):
     if spec.needs_mempool:
         _skip_if_no_mempool()
     g = GraphDefinition()
-    node, expected_attrs = spec.builder(g)
+    with xfail_on_graph_mempool_oom():
+        node, expected_attrs = spec.builder(g)
     return spec, g, node, expected_attrs
 
 
@@ -803,18 +807,20 @@ def test_alloc_zero_size_fails(sample_graphdef):
 def test_free_creates_dependency(sample_graphdef):
     """Free node depends on its predecessor."""
     _skip_if_no_mempool()
-    alloc = sample_graphdef.allocate(ALLOC_SIZE)
-    free = alloc.deallocate(alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        alloc = sample_graphdef.allocate(ALLOC_SIZE)
+        free = alloc.deallocate(alloc.dptr)
     assert alloc in free.pred
 
 
 def test_alloc_free_chain(sample_graphdef):
     """Alloc and free can be chained."""
     _skip_if_no_mempool()
-    a1 = sample_graphdef.allocate(ALLOC_SIZE)
-    a2 = a1.allocate(ALLOC_SIZE)
-    f2 = a2.deallocate(a2.dptr)
-    f1 = f2.deallocate(a1.dptr)
+    with xfail_on_graph_mempool_oom():
+        a1 = sample_graphdef.allocate(ALLOC_SIZE)
+        a2 = a1.allocate(ALLOC_SIZE)
+        f2 = a2.deallocate(a2.dptr)
+        f1 = f2.deallocate(a1.dptr)
     assert a1 in a2.pred
     assert a2 in f2.pred
     assert f2 in f1.pred
@@ -842,15 +848,17 @@ def test_alloc_device_option(sample_graphdef, device_spec):
     """Device can be specified as int or Device object."""
     _skip_if_no_mempool()
     device = Device()
-    node = sample_graphdef.allocate(ALLOC_SIZE, device=device_spec(device))
+    with xfail_on_graph_mempool_oom(device):
+        node = sample_graphdef.allocate(ALLOC_SIZE, device=device_spec(device))
     assert node.dptr != 0
 
 
 def test_alloc_peer_access(mempool_device_x2):
     """AllocNode.peer_access reflects requested peers."""
     d0, d1 = mempool_device_x2
     g = GraphDefinition()
-    node = g.allocate(ALLOC_SIZE, device=d0.device_id, peer_access=[d1.device_id])
+    with xfail_on_graph_mempool_oom(d0):
+        node = g.allocate(ALLOC_SIZE, device=d0.device_id, peer_access=[d1.device_id])
     assert d1.device_id in node.peer_access
 
 
@@ -863,8 +871,9 @@ def test_alloc_peer_access(mempool_device_x2):
 def test_join_merges_branches(sample_graphdef, num_branches):
     """join() with multiple branches creates correct dependencies."""
     _skip_if_no_mempool()
-    branches = [sample_graphdef.allocate(ALLOC_SIZE) for _ in range(num_branches)]
-    joined = sample_graphdef.join(*branches)
+    with xfail_on_graph_mempool_oom():
+        branches = [sample_graphdef.allocate(ALLOC_SIZE) for _ in range(num_branches)]
+        joined = sample_graphdef.join(*branches)
     assert isinstance(joined, EmptyNode)
     assert set(joined.pred) == set(branches)
 
@@ -956,8 +965,9 @@ def test_instantiate_empty_graph(sample_graphdef, inst_kwargs):
 def test_instantiate_with_nodes(sample_graphdef, inst_kwargs):
     """Graph with nodes can be instantiated."""
     _skip_if_no_mempool()
-    sample_graphdef.allocate(ALLOC_SIZE)
-    sample_graphdef.allocate(ALLOC_SIZE)
+    with xfail_on_graph_mempool_oom():
+        sample_graphdef.allocate(ALLOC_SIZE)
+        sample_graphdef.allocate(ALLOC_SIZE)
     graph = _instantiate(sample_graphdef, inst_kwargs)
     assert graph is not None
 
@@ -997,8 +1007,9 @@ def test_instantiate_and_execute_kernel(sample_graphdef, inst_kwargs):
 def test_instantiate_and_execute_alloc_free(sample_graphdef, inst_kwargs):
     """Graph with alloc/free can be executed."""
     _skip_if_no_mempool()
-    alloc = sample_graphdef.allocate(ALLOC_SIZE)
-    alloc.deallocate(alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        alloc = sample_graphdef.allocate(ALLOC_SIZE)
+        alloc.deallocate(alloc.dptr)
 
     stream = Device().create_stream()
     graph = _instantiate_and_upload(sample_graphdef, inst_kwargs, stream)
@@ -1010,9 +1021,10 @@ def test_instantiate_and_execute_alloc_free(sample_graphdef, inst_kwargs):
 def test_instantiate_and_execute_memset(sample_graphdef, inst_kwargs):
     """Graph with alloc/memset/free can be executed."""
     _skip_if_no_mempool()
-    alloc = sample_graphdef.allocate(ALLOC_SIZE)
-    ms = alloc.memset(alloc.dptr, 0xAB, ALLOC_SIZE)
-    ms.deallocate(alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        alloc = sample_graphdef.allocate(ALLOC_SIZE)
+        ms = alloc.memset(alloc.dptr, 0xAB, ALLOC_SIZE)
+        ms.deallocate(alloc.dptr)
 
     stream = Device().create_stream()
     graph = _instantiate_and_upload(sample_graphdef, inst_kwargs, stream)
@@ -1026,12 +1038,13 @@ def test_instantiate_and_execute_memcpy(sample_graphdef, inst_kwargs):
     _skip_if_no_mempool()
     import ctypes
 
-    src_alloc = sample_graphdef.allocate(ALLOC_SIZE)
-    dst_alloc = sample_graphdef.allocate(ALLOC_SIZE)
-    dep = sample_graphdef.join(src_alloc, dst_alloc)
-    ms = dep.memset(src_alloc.dptr, 0xAB, ALLOC_SIZE)
-    cp = ms.memcpy(dst_alloc.dptr, src_alloc.dptr, ALLOC_SIZE)
-    cp.deallocate(src_alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        src_alloc = sample_graphdef.allocate(ALLOC_SIZE)
+        dst_alloc = sample_graphdef.allocate(ALLOC_SIZE)
+        dep = sample_graphdef.join(src_alloc, dst_alloc)
+        ms = dep.memset(src_alloc.dptr, 0xAB, ALLOC_SIZE)
+        cp = ms.memcpy(dst_alloc.dptr, src_alloc.dptr, ALLOC_SIZE)
+        cp.deallocate(src_alloc.dptr)
 
     stream = Device().create_stream()
     graph = _instantiate_and_upload(sample_graphdef, inst_kwargs, stream)
@@ -1166,11 +1179,12 @@ def test_instantiate_and_execute_if_then(sample_graphdef):
     set_handle = mod.get_kernel("set_handle")
     add_one = mod.get_kernel("add_one")
 
-    alloc = sample_graphdef.allocate(ctypes.sizeof(ctypes.c_int))
-    ms = alloc.memset(alloc.dptr, 0, ctypes.sizeof(ctypes.c_int))
-    setter = ms.launch(LaunchConfig(grid=1, block=1), set_handle, condition, 1)
-    if_node = setter.if_then(condition)
-    if_node.then.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        alloc = sample_graphdef.allocate(ctypes.sizeof(ctypes.c_int))
+        ms = alloc.memset(alloc.dptr, 0, ctypes.sizeof(ctypes.c_int))
+        setter = ms.launch(LaunchConfig(grid=1, block=1), set_handle, condition, 1)
+        if_node = setter.if_then(condition)
+        if_node.then.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
 
     graph = sample_graphdef.instantiate()
     stream = Device().create_stream()
@@ -1198,13 +1212,14 @@ def test_instantiate_and_execute_if_else(sample_graphdef):
     set_handle = mod.get_kernel("set_handle")
     add_one = mod.get_kernel("add_one")
 
-    alloc = sample_graphdef.allocate(ctypes.sizeof(ctypes.c_int))
-    ms = alloc.memset(alloc.dptr, 0, ctypes.sizeof(ctypes.c_int))
-    setter = ms.launch(LaunchConfig(grid=1, block=1), set_handle, condition, 0)
-    ie_node = setter.if_else(condition)
-    ie_node.then.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
-    n1 = ie_node.else_.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
-    n1.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        alloc = sample_graphdef.allocate(ctypes.sizeof(ctypes.c_int))
+        ms = alloc.memset(alloc.dptr, 0, ctypes.sizeof(ctypes.c_int))
+        setter = ms.launch(LaunchConfig(grid=1, block=1), set_handle, condition, 0)
+        ie_node = setter.if_else(condition)
+        ie_node.then.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
+        n1 = ie_node.else_.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
+        n1.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
 
     graph = sample_graphdef.instantiate()
     stream = Device().create_stream()
@@ -1232,12 +1247,13 @@ def test_instantiate_and_execute_switch(sample_graphdef):
     set_handle = mod.get_kernel("set_handle")
     add_one = mod.get_kernel("add_one")
 
-    alloc = sample_graphdef.allocate(ctypes.sizeof(ctypes.c_int))
-    ms = alloc.memset(alloc.dptr, 0, ctypes.sizeof(ctypes.c_int))
-    setter = ms.launch(LaunchConfig(grid=1, block=1), set_handle, condition, 2)
-    sw_node = setter.switch(condition, 4)
-    for branch in sw_node.branches:
-        branch.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
+    with xfail_on_graph_mempool_oom():
+        alloc = sample_graphdef.allocate(ctypes.sizeof(ctypes.c_int))
+        ms = alloc.memset(alloc.dptr, 0, ctypes.sizeof(ctypes.c_int))
+        setter = ms.launch(LaunchConfig(grid=1, block=1), set_handle, condition, 2)
+        sw_node = setter.switch(condition, 4)
+        for branch in sw_node.branches:
+            branch.launch(LaunchConfig(grid=1, block=1), add_one, alloc.dptr)
 
     graph = sample_graphdef.instantiate()
     stream = Device().create_stream()
@@ -1272,7 +1288,8 @@ def test_conditional_node_type_preserved_by_nodes(sample_graphdef):
 def test_debug_dot_print_creates_file(sample_graphdef, dot_file):
     """debug_dot_print writes a DOT file."""
     _skip_if_no_mempool()
-    sample_graphdef.allocate(ALLOC_SIZE)
+    with xfail_on_graph_mempool_oom():
+        sample_graphdef.allocate(ALLOC_SIZE)
     sample_graphdef.debug_dot_print(str(dot_file))
     assert dot_file.exists()
     content = dot_file.read_text()
@@ -1282,7 +1299,8 @@ def test_debug_dot_print_creates_file(sample_graphdef, dot_file):
 def test_debug_dot_print_with_options(sample_graphdef, dot_file):
     """debug_dot_print accepts GraphDebugPrintOptions."""
     _skip_if_no_mempool()
-    sample_graphdef.allocate(ALLOC_SIZE)
+    with xfail_on_graph_mempool_oom():
+        sample_graphdef.allocate(ALLOC_SIZE)
     options = GraphDebugPrintOptions(verbose=True, handles=True)
     sample_graphdef.debug_dot_print(str(dot_file), options)
     assert dot_file.exists()
@@ -1291,6 +1309,7 @@ def test_debug_dot_print_with_options(sample_graphdef, dot_file):
 def test_debug_dot_print_invalid_options(sample_graphdef, dot_file):
     """debug_dot_print rejects invalid options type."""
     _skip_if_no_mempool()
-    sample_graphdef.allocate(ALLOC_SIZE)
+    with xfail_on_graph_mempool_oom():
+        sample_graphdef.allocate(ALLOC_SIZE)
     with pytest.raises(TypeError, match="options must be a GraphDebugPrintOptions"):
         sample_graphdef.debug_dot_print(str(dot_file), "invalid")