UPSTREAM PR #18413: ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON

[loci-dev]

Mirrored from ggml-org/llama.cpp#18413

Fixed compilation error when building with GGML_NATIVE=ON and CUDA Toolkit but without a GPU (e.g. docker build) where CMAKE_CUDA_ARCHITECTURES_NATIVE will be No CUDA Devices found., causing nvcc error Unsupported gpu architecture 'compute_No CUDA Devices found.'.

Also if CMAKE_CUDA_ARCHITECTURES is explicitly set by the user, maybe it's better to use it even when GGML_NATIVE=ON. Which is also consistent with the behavior before #18361.

Additionally moves the logging of CMAKE_CUDA_ARCHITECTURES after the replacement.

Possibly related to #18398.

[loci-review]

Explore the complete analysis inside the Version Insights

Perfect! I've generated the summary report for your project. Here's what the analysis shows:

Summary Report for llama.cpp PR #722

Key Findings:

• ✅ No significant performance impact detected
• All modified functions show less than 2% change in both response time and throughput
• The pull request is performance-neutral and safe to merge from a performance perspective

Project Information:

• Repository: auroralabs-loci/llama.cpp
• Pull Request: UPSTREAM PR #18413: ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON #722
• Comparing base version b5e08f92... against target version 2e238df7...

The analysis indicates that the changes in this PR maintain the existing performance profile of the application without introducing any performance regressions.