[BUG]: Tab completion with DeviceMemoryResource is broken #2053

Activity

1. leofang

   added this to the cuda.core v1.0.0 milestone on May 8, 2026
2. leofang

   added

   bugSomething isn't workingSomething isn't working

   P0High priority - Must do!High priority - Must do!

   cuda.coreEverything related to the cuda.core moduleEverything related to the cuda.core module

   on May 8, 2026
3. leofang commented on May 8, 2026

   

   leofang

   on May 8, 2026

   MemberAuthor

   More actions

   Tab completion for cuda.core.system.Device is broken too.
4. leofang

   added

   triageNeeds the team's attentionNeeds the team's attention

   on May 8, 2026
5. leofang commented on May 8, 2026

   

   leofang

   on May 8, 2026

   Last edited by leofang

   MemberAuthor

   More actions

   Root cause analysis

   In CPython 3.12+, rlcompleter.Completer.attr_matches() has a guard to skip evaluating Python properties (introduced by python/cpython#88915 / python/cpython#27401, for bpo-44752):

   if isinstance(getattr(type(thisobject), word, None), property):
       matches.append(match)   # skip getattr(obj, word)
       continue
   if (value := getattr(thisobject, word, None)) is not None:
       ...

   However, Cython cdef class properties generate getset_descriptor objects, not Python property objects. The isinstance(..., property) guard does not fire for them. When getattr(obj, word, None) is called, it only catches AttributeError — any other exception propagates uncaught, crashing attr_matches() entirely and producing zero completions.

   This was flagged as a concern before the naming audit landed — see #1945 (comment):

   [methods] are side-effect free with getattr/hasattr (properties are not)

   The offending commit (8b55a6c) converted get_allocation_handle() from a method to a @property. As a method, getattr(mr, 'get_allocation_handle') returned the bound method object (no side effects). As a property, getattr(mr, 'allocation_handle') invokes the getter, which raises RuntimeError for non-IPC resources — and crashes the completer.

   Note

   This is a known upstream CPython issue: python/cpython#112821 ("Reconsider BPO 44752 changes in rlcompleter since exceptions in descriptors cause auto completion to fail"). Filed Dec 2023, still open.

   Impact audit

   I audited every Cython @property in cuda/core/ that can raise a non-AttributeError exception. Only Cython cdef class properties (getset_descriptor) are affected — pure Python property descriptors are handled correctly by the rlcompleter guard.

   Tab completion is effectively broken for all classes listed below. The triggering conditions are not edge cases — they are the default or common states for these objects (non-IPC is the default for memory resources and events; NVML queries fail on many GPU models; contexts get closed during normal lifecycle).

   IPC-related properties (non-IPC is the default — affects most users):

   Class	Property	Exception	Condition
   DeviceMemoryResource	allocation_handle	RuntimeError	non-IPC MR (the default)
   PinnedMemoryResource	allocation_handle	RuntimeError	non-IPC MR (the default)
   Event	ipc_descriptor	RuntimeError	non-IPC event (the default)
   Buffer	ipc_descriptor	RuntimeError	buffer from non-IPC MR

   NVML-backed properties (raises on GPUs that don't support the query):

   Class	Property	Exception	Condition
   system.Device	attributes	NotSupportedError	GPU doesn't support NVML query
   system.Device	is_auto_boosted_clocks_enabled	NotSupportedError	GPU doesn't support feature
   system.Device	is_c2c_enabled	NotSupportedError	GPU doesn't support C2C
   system.Device	numa_node_id	NotSupportedError	GPU doesn't expose NUMA node

   Other properties:

   Class	Property	Exception	Condition
   Context	resources	RuntimeError	closed context
   DeviceResources	workqueue	RuntimeError	hardware-dependent
   GraphBuilder	is_building	NotImplementedError	unknown capture status
   system.Device	current_clock_event_reasons	ValueError	unknown clock event bit
   system.Device	supported_clock_event_reasons	ValueError	unknown clock event bit

   Not affected (abstract base, never instantiated directly):
   MemoryResource.{is_device_accessible, is_host_accessible, device_id} — TypeError, but only hit on bare MemoryResource instances.

   Not affected (pure Python class, rlcompleter guard works):
   LegacyPinnedMemoryResource.device_id — RuntimeError, but isinstance(desc, property) is True, so rlcompleter skips it.

   Possible solutions

   1. Upstream CPython fix (Reconsider gh-88915 changes rlcompleter since exceptions in descriptors cause auto completion to fail python/cpython#112821): Extend the rlcompleter guard to handle getset_descriptor (or restore the old try/except Exception around getattr). This would fix all Cython extension types, not just cuda.core.

   2. cuda-python workaround — raise AttributeError instead:
      getattr(obj, name, None) catches AttributeError, so changing raise RuntimeError(...) to raise AttributeError(...) in the affected property getters is a minimal one-line-per-property fix. Downside: semantically misleading (the attribute exists, it's just unusable in the current state), and CPython appends a "Did you mean: ...?" suggestion that can be confusing.

   3. cuda-python workaround — return None instead of raising:
      Properties like allocation_handle and ipc_descriptor could return None when the feature is not enabled. Callers already need to check is_ipc_enabled first.

   4. cuda-python workaround — revert to methods:
      Keep get_allocation_handle() / get_ipc_descriptor() as methods. Methods are never evaluated by getattr() during tab completion.

   5. cuda-python workaround — catch NVML errors in system.Device:
      Wrap NVML-backed properties in try/except and return a sentinel (e.g., None or a "not supported" marker) instead of letting NotSupportedError propagate.

   -- Leo's bot
6. mdboom commented on May 8, 2026

   

   mdboom

   on May 8, 2026

   Last edited by mdboom

   Contributor

   More actions

   Confirming that tab completion is /not/ broken in IPython:

   

   It is also /not/ broken in Jupyter:

   

   EDIT: And everything is available in this list if you scroll down, just not visible in the screenshot.

   So it seems confined to vanilla Python REPL tab completion specifically (which one could argue is the least important).

   🎉React with 🎉3Reacted by Keith Kraus, Oleksandr Pavlyk and Leo Fang
7. mdboom commented on May 8, 2026

   

   mdboom

   on May 8, 2026

   Contributor

   More actions

   Confirmed this affects Python 3.10 -> 3.14, so it's not just the "new" REPL introduced in 3.13.
8. mdboom

   self-assigned this

   on May 8, 2026
9. mdboom commented on May 8, 2026

   

   mdboom

   on May 8, 2026

   Contributor

   More actions

   My assessment of the agent's suggestions:

   Upstream CPython fix (python/cpython#112821): Extend the rlcompleter guard to handle getset_descriptor (or restore the old try/except Exception around getattr). This would fix all Cython extension types, not just cuda.core.

   This is a non-starter. Even if we got the bugfix in, it would be ~5 years until we could rely on it.

   cuda-python workaround — raise AttributeError instead:
   getattr(obj, name, None) catches AttributeError, so changing raise RuntimeError(...) to raise AttributeError(...) in the affected property getters is a minimal one-line-per-property fix. Downside: semantically misleading (the attribute exists, it's just unusable in the current state), and CPython appends a "Did you mean: ...?" suggestion that can be confusing.

   This would change the semantics of the property, and the "Did you mean" would be incredibly misleading to users. They didn't misspell -- they used a feature that semantically should raise some sort of runtime error. It should be fine for us to design that way.

   cuda-python workaround — return None instead of raising:
   Properties like allocation_handle and ipc_descriptor could return None when the feature is not enabled. Callers already need to check is_ipc_enabled first.

   This places the burden of error checking on our users. It's deeply un-Pythonic to require error code checks. We are even trying to move away from that in cuda_bindings.

   cuda-python workaround — revert to methods:
   Keep get_allocation_handle() / get_ipc_descriptor() as methods. Methods are never evaluated by getattr() during tab completion.

   This is probably the most reasonable thing to do on this list. If we went this way, we should add some sort of a test harness that makes sure we stick to the convention that "any property that can raise an exception must be a getter instead", because we will forget. It seems frustrating to design an API around the limitations of tab completion, though, so I'm still not a huge fan.

   cuda-python workaround — catch NVML errors in system.Device:
   Wrap NVML-backed properties in try/except and return a sentinel (e.g., None or a "not supported" marker) instead of letting NotSupportedError propagate.

   This is basically the same as #3 above.
10. leofang

    linked a pull request that will close this issueFix tab completion #2055on May 12, 2026
11. leofang

    closed this as completedon May 12, 2026