diff --git a/cuda_core/cuda/core/_cpp/rt/api.hpp b/cuda_core/cuda/core/_cpp/rt/api.hpp
index dabaac2a5c8..4abf85bd5dc 100644
--- a/cuda_core/cuda/core/_cpp/rt/api.hpp
+++ b/cuda_core/cuda/core/_cpp/rt/api.hpp
@@ -243,7 +243,8 @@ StreamHandle deallocation_stream(const DevicePtrHandle &h) noexcept;
 // Set the deallocation stream for a device pointer handle.
 // Returns CUDA_ERROR_INVALID_CONTEXT when a default-stream token cannot be
 // bound because no CUDA context is current.
-CUresult set_deallocation_stream(const DevicePtrHandle &h, const StreamHandle &h_stream) noexcept;
+// May throw std::bad_alloc.
+CUresult set_deallocation_stream(const DevicePtrHandle &h, const StreamHandle &h_stream);
 
 // ============================================================================
 // Virtual memory management (VMM_DESIGN.md)
diff --git a/cuda_core/cuda/core/_cpp/rt/internal.hpp b/cuda_core/cuda/core/_cpp/rt/internal.hpp
index 382c4901f11..9b0ef9d63aa 100644
--- a/cuda_core/cuda/core/_cpp/rt/internal.hpp
+++ b/cuda_core/cuda/core/_cpp/rt/internal.hpp
@@ -60,8 +60,9 @@ struct DeallocationStream {
 
 // Implemented in stream.cpp
 ContextHandle deallocation_context(const DeallocationStream &stream) noexcept;
-// Implemented in stream.cpp
-bool make_deallocation_stream(const StreamHandle &h, DeallocationStream &out) noexcept;
+// Implemented in stream.cpp. Not noexcept: binding a default-stream token
+// allocates. May throw std::bad_alloc.
+bool make_deallocation_stream(const StreamHandle &h, DeallocationStream &out);
 
 // Implemented in virtual_memory.cpp. Synchronize the stream a VMM buffer
 // recorded before its mappings are released. Skips the sync, with a report
diff --git a/cuda_core/cuda/core/_cpp/rt/memory.cpp b/cuda_core/cuda/core/_cpp/rt/memory.cpp
index f280aeedc5b..a80ffd3a2a1 100644
--- a/cuda_core/cuda/core/_cpp/rt/memory.cpp
+++ b/cuda_core/cuda/core/_cpp/rt/memory.cpp
@@ -193,7 +193,7 @@ deallocation_stream(const DevicePtrHandle &h) noexcept
 
 // Replace the stream that orders a device pointer's deallocation.
 CUresult
-set_deallocation_stream(const DevicePtrHandle &h, const StreamHandle &h_stream) noexcept
+set_deallocation_stream(const DevicePtrHandle &h, const StreamHandle &h_stream)
 {
     if (!h) {
         return CUDA_ERROR_INVALID_VALUE;
diff --git a/cuda_core/cuda/core/_cpp/rt/stream.cpp b/cuda_core/cuda/core/_cpp/rt/stream.cpp
index eba71ca5be8..917106a42c4 100644
--- a/cuda_core/cuda/core/_cpp/rt/stream.cpp
+++ b/cuda_core/cuda/core/_cpp/rt/stream.cpp
@@ -215,9 +215,11 @@ create_context_bound_legacy_stream(const ContextHandle &h_context)
 namespace detail {
 // Real streams are copied unchanged. Default-stream tokens without an embedded
 // context are bound to the current context. Returns false (and sets err) when a
-// default-stream token cannot be bound because no context is current.
+// default-stream token cannot be bound because no context is current. Binding
+// a default-stream token allocates (the bound StreamBox and the current-context
+// handle), so this may throw std::bad_alloc and must not be noexcept.
 bool
-make_deallocation_stream(const StreamHandle &h, DeallocationStream &out) noexcept
+make_deallocation_stream(const StreamHandle &h, DeallocationStream &out)
 {
     out = {};
     if (!h) {
diff --git a/cuda_core/cuda/core/_memory/_buffer.pyx b/cuda_core/cuda/core/_memory/_buffer.pyx
index 23ab10c06e9..52fba667886 100644
--- a/cuda_core/cuda/core/_memory/_buffer.pyx
+++ b/cuda_core/cuda/core/_memory/_buffer.pyx
@@ -316,7 +316,13 @@ cdef class Buffer:
             record_stream = mr.is_device_accessible
             self._h_ptr = deviceptr_create_with_mr(c_ptr, size, mr)
             if record_stream:
-                _ds_status = set_deallocation_stream(self._h_ptr, s._h_stream)
+                try:
+                    _ds_status = set_deallocation_stream(self._h_ptr, s._h_stream)
+                except MemoryError:
+                    # Reset for the same reason as below, inside the handler so the
+                    # destructor cannot swallow the MemoryError that is propagating.
+                    self._h_ptr.reset()
+                    raise
                 if _ds_status != cydriver.CUresult.CUDA_SUCCESS:
                     # Reset before raising: the DevicePtrHandle destructor would otherwise
                     # invoke _mr_dealloc_callback, which catches any inner exception and
diff --git a/cuda_core/cuda/core/_rt.pxd b/cuda_core/cuda/core/_rt.pxd
index 4d5e63cb2bc..60617192c4b 100644
--- a/cuda_core/cuda/core/_rt.pxd
+++ b/cuda_core/cuda/core/_rt.pxd
@@ -295,7 +295,7 @@ cdef void register_mr_dealloc_callback(MRDeallocCallback cb) noexcept
 cdef DevicePtrHandle deviceptr_import_ipc(
     const MemoryPoolHandle& h_pool, const void* export_data, const StreamHandle& h_stream) except+ nogil
 cdef StreamHandle deallocation_stream(const DevicePtrHandle& h) noexcept nogil
-cdef cydriver.CUresult set_deallocation_stream(const DevicePtrHandle& h, const StreamHandle& h_stream) noexcept nogil
+cdef cydriver.CUresult set_deallocation_stream(const DevicePtrHandle& h, const StreamHandle& h_stream) except+ nogil
 
 # Virtual memory management (VMM_DESIGN.md)
 cdef MemAllocationHandle create_mem_allocation_handle(
diff --git a/cuda_core/cuda/core/_rt.pyx b/cuda_core/cuda/core/_rt.pyx
index 49523bbce91..ee7db861d6e 100644
--- a/cuda_core/cuda/core/_rt.pyx
+++ b/cuda_core/cuda/core/_rt.pyx
@@ -175,7 +175,7 @@ cdef extern from "_cpp/rt/rt.hpp" namespace "cuda_core::rt":
     StreamHandle deallocation_stream "cuda_core::rt::deallocation_stream" (
         const DevicePtrHandle& h) noexcept nogil
     cydriver.CUresult set_deallocation_stream "cuda_core::rt::set_deallocation_stream" (
-        const DevicePtrHandle& h, const StreamHandle& h_stream) noexcept nogil
+        const DevicePtrHandle& h, const StreamHandle& h_stream) except+ nogil
 
     # Virtual memory management (VMM_DESIGN.md)
     MemAllocationHandle create_mem_allocation_handle "cuda_core::rt::create_mem_allocation_handle" (
diff --git a/cuda_core/docs/source/release/1.3.0-notes.rst b/cuda_core/docs/source/release/1.3.0-notes.rst
index 7a0d6cb0e01..e17192853d8 100644
--- a/cuda_core/docs/source/release/1.3.0-notes.rst
+++ b/cuda_core/docs/source/release/1.3.0-notes.rst
@@ -163,6 +163,10 @@ Fixes and enhancements
 - Fixed a ``TypeError`` in :meth:`Program.compile` when retrying after PCH
   heap exhaustion.
 
+- :meth:`Buffer.set_deallocation_stream` and :meth:`Buffer.from_handle` now raise
+  :class:`MemoryError` when the host runs out of memory while binding a default
+  stream, instead of aborting the process with ``std::terminate``.
+
 - :attr:`system.Device.uuid_without_prefix` preserves UUIDs that NVML already
   reports without a prefix, such as on Orin. This also allows
   :meth:`system.Device.to_cuda_device` to match those devices correctly.