fix(moore): select MCCL bfloat16 architecture

[voltjia]

Summary

Moore MCCL exposes mcclBfloat16 only when MARCH_TYPE describes a capable target architecture. InfiniCCL previously passed neither MARCH_TYPE nor an offload architecture to mcc, so BF16 collectives returned infinicclNotSupported even on S5000 hardware where MCCL supports them.

This PR adds explicit and auto-detected MUSA architecture configuration, passes the matching compiler flags, and keeps mixed-architecture builds fail-closed. The architecture handling is organized under reusable CMake helpers so future SDKs with macro-defined capability surfaces can follow the same pattern without duplicating the common logic.

Changes

• Architecture selection

  • Add MUSA_ARCHITECTURES, accepting forms such as 31, mp_31, and 3.1.
  • Prefer the explicit cache value, then TORCH_MUSA_ARCH_LIST, then native detection through musaGetDeviceProperties.
  • Add MUSA architecture resolution in cmake/architecture/MusaArchitecture.cmake, keeping the root Moore CMake block focused on SDK and library discovery.
  • Add generic macro-architecture helpers in cmake/architecture/MacroArchitecture.cmake for normalization, conservative macro selection, compiler-option generation, and target-scoped application.

• MCCL capability mapping

  • Pass one --offload-arch=mp_* flag per configured target.
  • Define MARCH_TYPE from the least capable target because mccl.h exposes one global data-type table; this prevents a mixed fat binary from advertising BF16 on an unsupported target.
  • Apply the same architecture contract to the shared library and every Moore example target.

Platform and Backend Affected

Platform

• CPU
• NVIDIA GPU
• Iluvatar GPU
• MetaX GPU
• Moore Threads GPU
• Cambricon MLU
• HYGON DCU

Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

Performance Impact

• No performance impact
• Performance improved
• Performance regression possible

N/A. This changes build-time target selection and enables an existing MCCL data type.

Known Issues & Future Work

• Native architecture detection requires a visible MUSA device. Cross-builds must set MUSA_ARCHITECTURES or TORCH_MUSA_ARCH_LIST explicitly.

Test Results

Test Involved Platform

• CPU
• NVIDIA GPU
• Iluvatar GPU
• MetaX GPU
• Moore Threads GPU
• Cambricon MLU
• HYGON DCU

Test Involved Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

Validated on Moore Threads S5000 with MUSA 4.3.6:

Explicit fat target 31;22:
  MUSA_ARCHITECTURES=31;22
  MARCH_TYPE=220
  compile flags: --offload-arch=mp_31 --offload-arch=mp_22 -DMARCH_TYPE=220
  InfiniCCL build: PASS
  CTest: 2/2 passed

Native detection:
  MUSA_ARCHITECTURES=31
  MARCH_TYPE=310
  compile flags: --offload-arch=mp_31 -DMARCH_TYPE=310
  InfiniCCL build: PASS
  BUILD_EXAMPLES=ON full build: PASS
  example compile flags: --offload-arch=mp_31 -DMARCH_TYPE=310
  CTest: 2/2 passed

Two-device BF16 AllReduce:
  input ranks: 1.0, 2.0
  output on both ranks: 3.0
  result: PASS

This commit was also combined with the open AllGather/Send/Recv stack at e8ccc0cb for InfiniLM validation. Qwen3-32B TP4 and Baichuan2-7B TP2 paged FlashAttention graph inference both passed after BF16 AllReduce became available.

Moore MCCL + OpenMPI (single node):
ccl_mpi_hybrid_all_reduce.log

Moore MCCL (single node):
ccl_all_reduce.log

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows the repository convention.
• Each commit message follows Conventional Commits.
• This is a single squashable commit.
• The branch is based cleanly on current master with no merge commit.
• No fixup, squash, or work-in-progress commits remain.

Scope and Design

• Changes are minimal and Moore-specific.
• No dead code, debug output, or unowned TODO remains.
• No unrelated formatting churn was introduced.
• The new build option is documented in its CMake cache description and errors.

General Code Hygiene

• The code is self-explanatory; comments were added only where the intent or rationale is non-obvious (CONTRIBUTING.md §Code/General).
• Every modified or added file ends with a single trailing newline (CONTRIBUTING.md §Code/General).
• No trailing whitespace, inconsistent indentation, or mixed formatting styles remain.
• Identifiers referenced in comments or error messages are wrapped in Markdown backticks (e.g. the `AllReduce` implementation) (CONTRIBUTING.md §Code/General).
• All comments and error messages are in English (CONTRIBUTING.md §Code/General).
• Comments and error messages are complete sentences — capitalized first letter, terminal punctuation — unless the language/framework convention says otherwise (CONTRIBUTING.md §Code/General; §Python).

C++ Specific

N/A. No repository C++ source file is modified; the native detector is generated by CMake during configuration.

Python Specific

N/A. No Python file is modified.

Testing

• The Moore/MCCL library, CMake tests, and a two-device BF16 AllReduce were built and run successfully.

Build, CI, and Tooling

• Existing Moore and MCCL detection remains unchanged; architecture detection runs only when WITH_MOORE=ON.
• The branch is ready for repository CI.

Documentation

• The user-visible cache option and cross-build requirement are documented inline.
• No breaking change is introduced.

Security and Safety

• No secret, internal URL, customer data, or personal hardware identifier is committed.
• No third-party source code is added.
• No pointer arithmetic, memory access, or bounds-sensitive runtime code is changed.