[baominghelly]

Summary

• Replace the four ambiguous Hygon CUDA function aliases (GetDevice, GetDeviceCount, DeviceSynchronize, and MemcpyAsync) with forwarding lambdas, matching the existing wrappers in src/native/cuda/hygon/runtime_.h.
• Compile and run the existing Hygon native runtime suite through an additional .cu translation unit, retaining the ordinary C++ test.

Motivation

In the DTK 25.04.1 CI image, InfiniRT builds successfully as C++, but downstream InfiniOps device compilation fails when it includes the installed Hygon runtime header:

variable 'GetDevice' with type 'const auto' has incompatible initializer of type '<overloaded function type>'

The same diagnostic occurs for all four aliases. Calling the CUDA APIs inside forwarding wrappers allows overload resolution at the call site instead of trying to deduce a function pointer from an overload set.

The baseline failure was reproduced with a minimal CUDA translation unit before validating the fix. Original downstream failure.

Type of Change

• feat - new feature / new backend capability / new public API
• fix - bug fix
• perf - performance improvement without behavior change
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• Hygon (WITH_HYGON)
• MetaX (WITH_METAX)
• Moore (WITH_MOORE)
• Cambricon (WITH_CAMBRICON)
• Ascend (WITH_ASCEND)
• Build system / CMake / generated headers
• Public headers / installed consumer API
• Documentation only

The CMake addition is restricted to WITH_HYGON. Generated and installed Hygon detail headers inherit the source change; the header generator itself is unchanged.

Smoke Build and Test Result

Validated commit: 46d411ef37f7c8a5d599ceca78547eac0bcbbfd4.

Environment: BW200 / gfx936, image image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.4.1-ubuntu22.04-dtk25.04.1-py3.10, DTK CUDA facade 12.6.77. The temporary container selected its existing CUDA 12 toolkit and exported CUDA_PATH, CUDA_HOME, PATH, and its runtime library directory in LD_LIBRARY_PATH. One idle GPU was selected automatically and held with the existing host GPU lease; the container verified that exactly one device was visible.

cmake -S /task/infinirt-fixed -B /task/fixed-build \
  -DWITH_CPU=ON -DWITH_HYGON=ON -DINFINI_RT_BUILD_TESTING=ON \
  -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/task/infinirt-prefix
cmake --build /task/fixed-build -j 8
ctest --test-dir /task/fixed-build --output-on-failure
cmake --install /task/fixed-build
nvcc -std=c++17 -arch=gfx936 -I/task/infinirt-prefix/include \
  -c /task/header-consumer.cu -o /task/fixed-header.o

100% tests passed, 0 tests failed out of 14
Includes test_hygon_runtime, test_hygon_cuda_runtime, test_hygon_graph,
test_runtime_dispatch, and test_install_consumer.
Patched installed header compiled under DTK nvcc.

Downstream InfiniOps build and smoke:

python -m pip install '.[dev]' --no-build-isolation \
  --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON \
  --config-settings=cmake.define.INFINI_RT_ROOT=/task/infinirt-prefix
pytest tests/ -m smoke -n 4 -v --tb=short --junitxml=/task/test-results.xml

InfiniOps wheel built and installed successfully.
161 passed, 73 skipped, 2 warnings in 16.36s

A fresh GitHub runner acceptance run pinned this exact InfiniRT commit and independently passed build/install and the same smoke suite: 161 passed, 73 skipped, 0 failed. JUnit and GPU lease artifact.

The existing smoke suite determines the skips for unavailable implementations and platform-specific requirements. No test selection, tolerance, or idle-GPU threshold was relaxed. The two Torch warnings concern unavailable memory-efficient attention; the corresponding tests passed.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	No	N/A - not affected	CPU tests also passed in the combined CPU+Hygon configuration.
NVIDIA	No	N/A - not affected	Backend code and test configuration unchanged.
Iluvatar	No	N/A - not affected	Backend code and test configuration unchanged.
Hygon	Yes	smoke passed	Complete configured InfiniRT CTest suite: 14/14; downstream InfiniOps smoke: 161 passed, 73 skipped. Full downstream operator suite not run.
MetaX	No	N/A - not affected	Backend code and test configuration unchanged.
Moore	No	N/A - not affected	Backend code and test configuration unchanged.
Cambricon	No	N/A - not affected	Backend code and test configuration unchanged.
Ascend	No	N/A - not affected	Backend code and test configuration unchanged.

Benchmark / Performance Impact

N/A. This fixes device-compilation compatibility; no performance claim.

Notes for Reviewers

• The additional CUDA test reuses test_native_runtime.cc and the existing Hygon test definitions so host and device compiler coverage exercise the same runtime behavior.
• On a separate BW1100 host with DTK 26.04 / Clang 17, the unpatched baseline header compiled with that host's DTK compiler after supplying its C++ standard-library include paths. The overload failure was not reproduced there. This comparison does not isolate whether compiler or compatibility-header changes are responsible, and is not a full validation of the patched suite on DTK 26.04.
• Clang Format and Ruff passed for the fix commit.