UPSTREAM PR #18443: ggml-cuda: (cmake) expand "native" into concrete architectures

[loci-dev]

Mirrored from ggml-org/llama.cpp#18443

Fixes #18430 and possibly #18434.

#18413 caused error on blackwell local build because native will be treated as 120-real but is not captured by the regex.

Tested with or without docker (with a blackwell GPU). Seems OK.

[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:

Key Findings:

✅ No Performance Issues Detected

The performance analysis of Pull Request #735 in the llama.cpp repository shows that:

• No modified functions have performance changes greater than 2% threshold
• Both Response Time and Throughput Time metrics remain stable
• All changes are within acceptable performance variance

Recommendation:

This pull request appears to be safe to merge from a performance perspective. The changes introduced do not cause any significant performance regressions, with all modified functions maintaining performance within 2% of the baseline version.

This is a positive result indicating that the code changes maintain the performance characteristics of the existing codebase.