fix: make device header self-contained

[voltjia]

Summary

Make src/device.h self-contained by including <cstdint> directly and using std::uint8_t as the underlying type of MemorySpace.

This removes reliance on a transitive declaration of uint8_t. It does not add collective functionality or change runtime behavior, enum values, or supported platforms and backends.

Changes

• Header dependency

  • Include <cstdint> directly in src/device.h.

• Fixed-width type

  • Qualify the underlying type of MemorySpace as std::uint8_t.

Platform and Backend Affected

src/device.h is shared by every device implementation and the common communication paths, so all supported configurations are affected at compile time. Runtime behavior is unchanged.

Platform

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

Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

Performance Impact

• No performance impact
• Performance improved
• Performance regression possible

N/A. This change only makes a header dependency explicit and preserves the existing fixed-width enum representation.

Known Issues & Future Work

• No known issues are introduced by this change.
• Runtime validation for this PR is limited to NVIDIA GPU with NCCL. Other affected platform/backend combinations were not exercised.
• This PR does not add or expand collective, platform, or backend functionality.

Test Results

Environment:

• Container: accelerator-dev/nvidia:latest (NVIDIA PyTorch 25.12)
• Hardware: 2 x NVIDIA A100-SXM4-80GB
• Configuration: NVIDIA device with NCCL backend, two local ranks

Results:

• clang-format-16 --dry-run --Werror src/device.h: passed.
• ruff check .: passed.
• ruff format --check .: passed.
• Standalone C++17 compilation including only src/device.h: passed.
• scripts/run_examples.py for ccl/all_reduce: passed with two NVIDIA GPUs.
• Downstream InfiniLM Qwen3-0.6B inference: passed with tensor parallelism 1 and 2.

The applicable example was run as:

python scripts/run_examples.py \
    --config /workspace/ccl-cluster.yaml \
    --examples ccl/all_reduce \
    --log-path /workspace/ccl-example-logs \
    --timeout 600 \
    -- -g 2 -w 1 -p 1 -n 1048576

Generated ccl_all_reduce.log result:

[Main Process] Host: master | Target GPUs: 2

=== AllReduce Results ===
Correct: YES
Expect:  3.00
Actual:  3.00

=== Single-Node Threaded AllReduce Results ===
Data size: 1048576 floats (4 MB)
Time:           0.119 ms
Throughput:     32.71 GB/s (Bus BW)
Alg Bandwidth:  32.71 GB/s

[Main Process] All worker threads joined. InfiniCCL finalized safely.

Test Involved Platform

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

Test Involved Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz, with the type matching the PR title.
• Each commit message follows Conventional Commits.
• This small PR is a single squashable commit.
• The branch is rebased cleanly on top of the current master with no stray merge commits.
• No fixup!, squash!, or wip commits remain.

Scope and Design

• Changes are minimal and contain no unrelated modifications.
• No dead code, commented-out blocks, debug prints, or unowned TODO items were introduced.
• No unrelated formatting churn obscures the diff.
• N/A - Public API behavior, ABI, and enum values are unchanged.

General Code Hygiene

• The change is self-explanatory and needs no additional comment.
• The modified file ends with a single trailing newline.
• No trailing whitespace, inconsistent indentation, or mixed formatting styles remain.
• N/A - No comments or error messages were added or modified.
• N/A - No comments or error messages were added or modified.
• N/A - No comments or error messages were added or modified.

C++ Specific (if C++ files changed)

• Code follows the Google C++ Style Guide.
• clang-format version 16 has been run against the modified file; the diff is clean.
• N/A - No exception or error path was added or modified.
• N/A - No error or warning message was added or modified.
• N/A - No constructor or initializer list was added or modified.
• N/A - No class or function spacing was changed.
• N/A - No class member spacing was changed.
• N/A - No namespace layout was changed.

Python Specific (if Python files changed)

• N/A - No Python files changed.
• N/A - No Python files changed.
• N/A - No Python comments changed.
• N/A - No Python framework-specific messages changed.
• N/A - No Python function signatures changed.
• N/A - No Python control flow changed.
• N/A - No Python return statements changed.
• N/A - No Python docstrings changed.
• N/A - No Python type hints changed.

Testing

• The applicable example program was built and tested successfully with two NVIDIA GPUs and NCCL.

Build, CI, and Tooling

• N/A - No backend, device, or auto-detection logic changed.
• The clang-format.yml and ruff.yml checks pass locally.

Documentation

• N/A - No behavior, build flag, or developer workflow changed.
• N/A - This PR has no user-visible breaking change.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers were committed.
• N/A - No third-party code was added or modified.
• No unsafe pointer arithmetic, uninitialized read, or missing bounds check was introduced.