Summary

This PR adds end-to-end CUDA support for the nncase NTT path and native runtime. It introduces a CUDA target/module compiler, CUDA runtime module loading and execution support, CUDA-aware NTT runtime primitives, CUDA kernel tests, and a dedicated Linux CUDA CI job for running those tests separately from the regular CPU/macOS compiler jobs.

Motivation

• Enable nncase to compile NTT-generated kernels for CUDA and execute them through the native runtime, instead of stopping at code generation.
• Keep CPU and CUDA generated modules on compatible runtime/operator ABI boundaries while allowing CUDA-specific device entry points and launch behavior.
• Catch CUDA-specific regressions in CI without making ordinary Linux/macOS compiler jobs depend on GPU availability.
• Fix issues uncovered while enabling CUDA tests, including CUDA toolkit discovery, device-callable scalar helpers, generated module ABI handling, and reduce-axis normalization in the NTT vectorization/lowering path.

Implementation

• Added CUDA target plumbing in Nncase.Modules.NTT, including CUDATarget, CUDAModuleCompiler, target abstraction cleanup, and CUDA-aware C/CMake generation.
• Added native CUDA runtime support with CUDA runtime module/function classes, loader integration, runtime CMake wiring, and ENABLE_CUDA_RUNTIME gating.
• Added NTT CUDA runtime support for topology, remote tensors, distributed operations, vector ops, profiling, and CUDA runtime entry points.
• Updated NTT kernels and runtime utilities so generated code can compile for both CPU and CUDA, including device-callable scalar conversions/operators for half and related scalar types.
• Normalized negative reduce axes during NTT vectorization/lowering instead of IR construction, preserving IR semantics while fixing CUDA reduce vectorization cases.
• Added CUDA kernel test coverage through UnitTestCUDAKernels and enabled it in CI with a dedicated test-x86_64-linux-cuda job.
• Kept the regular compiler test job excluding UnitTestCUDAKernels so CPU-only Linux/macOS jobs remain independent from CUDA runtime availability.

Validation

• git diff --check
• YAML parsing for .github/workflows/compiler-build.yml
• dotnet build modules/Nncase.Modules.NTT/Nncase.Modules.NTT.csproj -c Release --no-restore
• Rebuilt the native runtime locally with clang, CUDA 12.8, and ENABLE_CUDA_RUNTIME=ON.
• Verified the installed native runtime exposes CUDA runtime support and links against CUDA runtime libraries.
• Verified generated CUDA modules compile locally with clang++ and CUDA 12.8 for the reduce/vectorization repro cases.
• dotnet test src/Nncase.Tests/Nncase.Tests.csproj -c Release --no-build --no-restore --filter "FullyQualifiedName~Nncase.Tests.TargetTest.UnitTestCUDAKernels.TestVectorizeReduce": 8/8 passed locally.
• Existing PR compiler/runtime/code-format checks passed before enabling the dedicated CUDA CI job; the latest CI run is rerunning with the new CUDA check included.

Limitations

• Full UnitTestCUDAKernels execution requires a Linux runner with an NVIDIA GPU, CUDA toolkit, nvcc, clang/clang++, and the labels self-hosted, linux, x64, cuda. GitHub-hosted CPU runners cannot execute these runtime tests.
• The dedicated CUDA CI job currently uses CUDA architecture 120, matching the local validation environment.
• The general compiler test job still intentionally excludes UnitTestCUDAKernels; CUDA runtime tests are expected to run only in the dedicated CUDA job.
• HuggingFace importer tests still depend on an available HF_HOME cache/configuration in CI and should be revisited separately from the CUDA runtime work.

Backlog

Future CUDA follow-up work is tracked in #2.