fix: add CPU memory space detection

[xingjiangjun]

Summary

Add the missing GetMemorySpace<Device::Type::kCpu> specialization in src/devices/cpu/device_.h.

The CPU device currently lacks this specialization, which can cause undefined reference errors when CPU execution paths use GetMemorySpace<kCpu>.

This change completes the CPU memory-space interface by returning MemorySpace::kHost.

Changes

• CPU Device
  • Implement the missing GetMemorySpace<kCpu> specialization.
  • Classify CPU memory as MemorySpace::kHost.
  • Keep the change limited to src/devices/cpu/device_.h, without modifying public APIs or communication algorithms.

Platform and Backend Affected

Platform

• CPU
• NVIDIA GPU
• Iluvatar GPU
• MetaX GPU
• Moore Threads GPU
• Cambricon MLU
• HYGON DCU
• T-Head PPU

Backend

• OpenMPI
• MPICH
• NCCL/RCCL
• MCCL
• CNCL

The implementation is device-level and is not inherently tied to a specific communication backend. OpenMPI is the validated configuration.

Performance Impact

• No performance impact
• Performance improved
• Performance regression possible

No performance impact is expected because this change provides a missing memory-space specialization without modifying existing communication algorithms.

Benchmark: N/A.

Known Issues & Future Work

• No functional regressions were observed in the tested CPU/OpenMPI configuration.
• Other backend and device configurations have not been validated.

Test Results

Test Environment

• Container: NVIDIA PyTorch 24.10
• Compiler: GCC 11.4
• MPI: OpenMPI 4.1.7
• Platform: CPU-only
• MPI ranks: 2 local ranks
• Base code: Current PR branch based on upstream master

Build configuration:

-DAUTO_DETECT_DEVICES=OFF
-DWITH_NVIDIA=OFF
-DWITH_OMPI=ON
-DBUILD_EXAMPLES=ON

Test Execution

Executed the official scripts/run_examples.py test suite using the CPU-only OpenMPI configuration.

Results:

• Total tests: 9
• Passed: 9
• Failed: 0
• Success rate: 100%

Passed examples:

• mpi/all_reduce
• mpi/all_gather
• mpi/reduce
• mpi/all_to_all
• mpi/reduce_scatter
• mpi/scatter
• mpi/broadcast
• mpi/gather
• mpi/send_recv

The Broadcast example also passed all three scenarios:

• Out-of-Place Broadcast
• In-Place Broadcast
• Legacy In-Place Bcast

All selected examples completed successfully.

Per-example test logs were generated by the official test harness and will be provided as an archive in a follow-up PR comment.

Test Involved Platform

• CPU
• NVIDIA GPU
• Iluvatar GPU
• MetaX GPU
• Moore Threads GPU
• Cambricon MLU
• HYGON DCU
• T-Head PPU

Test Involved Backend

• OpenMPI
• MPICH
• NCCL/RCCL
• MCCL
• CNCL

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows the fix/ naming convention.
• Commit message follows Conventional Commits.
• The PR contains a single focused commit.
• No unrelated merge commits are included.
• No fixup!, squash!, or wip commits remain.

Scope and Design

• Changes are minimal and limited to the CPU memory-space specialization.
• No unrelated code, debug output, or TODOs were introduced.
• No unrelated formatting changes were introduced.
• No public API changes were made.

General Code Hygiene

• The implementation is concise and self-explanatory.
• The trailing newline and whitespace conventions have not been independently checked by a formatting tool.
• The added code follows the surrounding indentation style.
• N/A: No new diagnostic messages or explanatory comments were introduced.

C++ Specific

• The implementation follows the surrounding C++ code structure and indentation conventions.
• clang-format-16 has not been run because it is unavailable in the current environment.
• No exceptions are thrown by the new implementation.
• No unsafe pointer arithmetic or memory operations were introduced.
• The added function is separated from the preceding declaration by one blank line.
• Namespace spacing follows the existing source layout.
• N/A: No constructors, class members, or diagnostic messages were added.

Python Specific

• N/A: No Python files were modified.

Testing

• The CPU-only OpenMPI build completed successfully.
• All nine selected MPI examples passed using two local CPU ranks.
• Broadcast passed all three supported scenarios in its example.
• A separate heterogeneous multi-node cluster test has not been performed.

Build, CI, and Tooling

• N/A: No new device or backend was introduced, so CMake auto-detection does not require changes.
• Local CI formatting workflows have not been executed.
• GitHub CI results are pending.

Documentation

• N/A: No new public API, build option, or user-facing workflow was introduced.
• N/A: No breaking changes were introduced.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers were added.
• N/A: No third-party code was introduced.
• No unsafe pointer arithmetic, uninitialized reads, or missing bounds checks were introduced.
  run_20261014_121207.zip