UPSTREAM PR #18426: Revert "ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATI…

[loci-dev]

Mirrored from ggml-org/llama.cpp#18426

…VE=ON (#18413)"

This reverts commit 4fd59e8.

The reason is that fix is only for the Docker builds without GPU and breaks all native builds otherwise. We also cannot allow users to specify 120-real and expect compilation to work at the moment. For Docker builds without GPU the correct way would be to build without GGML_NATIVE

[loci-review]

Explore the complete analysis inside the Version Insights

I've generated a summary report for your project. Here are the key findings:

Performance Summary Report

Project Details:

• Repository: llama.cpp (auroralabs-loci)
• Pull Request: UPSTREAM PR #18426: Revert "ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATI… #727
• Project ID: 2621b8c0-b5ce-11f0-b333-453f42058aa1
• Report ID: 3ea57dbe-1bf2-4717-9fa7-a6eb14c07018

Version Comparison:

• Base Version: b8001133-46bc-4b58-8e78-3e22fd732519
• Target Version: 6d193b1f-78b3-41a6-b2e9-b3986ba39cf1

---

Key Findings

Performance Impact: ✅ MINIMAL

The analysis shows that no modified functions were found with performance changes greater than 2% for both:

• Response Time - No significant changes detected
• Throughput Time - No significant changes detected

Summary

This pull request (#727) appears to have minimal to no performance impact on the llama.cpp codebase. All modified functions show performance variations within the 2% threshold, which is typically considered within normal variance and not a significant regression or improvement.

Recommendation: From a performance perspective, this change appears safe to merge as it does not introduce any notable performance regressions or improvements.