Build GPU conda package

[moazreyad]

Summary

This PR adds a GPU conda build workflow and updates the SINGA GPU build path for CUDA 12.2 / cuDNN 8.

Changes

• Add .github/workflows/conda-gpu-build.yaml

  • Runs on ubuntu-latest
  • Builds the GPU conda package with conda-build
  • Uses CUDA 12.2
  • Uses Python 3.8

• Update the conda GPU recipe

  • Pin CMake below 4 for compatibility with the current project CMake files
  • Align CUDA 12.2 build dependencies
  • Use GCC/G++ 11 for the CUDA 12 build
  • Add CUDA runtime/dev package pins needed by the build

• Update CUDA/cuDNN compatibility

  • Add missing Thrust header include
  • Support cuDNN 8 convolution algorithm selection APIs
  • Support cuDNN 8 RNN descriptor setup APIs

Validation

• Confirmed conda render succeeds.

• Confirmed conda-build tool/conda/singa succeeds in GitHub Actions.

• Downloaded and tested the built conda package locally.

• Verified the package can be installed from a local indexed conda channel with runtime dependencies resolved.

• Verified import smoke test:

  python -c "import singa; import singa.singa_wrap; print(singa.__version__)"
  

• Verified lightweight GPU examples locally:

  python examples/mlp/model.py --max-epoch 1
  python examples/mlp/native.py --max-epoch 1

Notes

GitHub-hosted ubuntu-latest runners do not provide NVIDIA GPUs, so the workflow currently validates that the GPU conda package builds successfully. Runtime GPU execution was tested locally on a machine with an NVIDIA GPU.