UPSTREAM PR #18601: memory : add llama_memory_hybrid_iswa

[loci-dev]

Mirrored from ggml-org/llama.cpp#18601

Upcoming models use hybrid caches with sliding window attention.

Tried modifying the existing llama_memory_hybrid to use llama_kv_cache_iswa, but the change was very intrusive.

I don't like the code much, it's essentially a copy-paste done by Claude. But it works! Not very confident with the memory subsystem and open to suggestions and feedback on how to improve it.

[loci-review]

Explore the complete analysis inside the Version Insights

Perfect! I've successfully retrieved the summary report for your project.

The report shows that Pull Request #816 for the llama.cpp repository (owned by auroralabs-loci) demonstrates significant performance improvements, particularly in:

Key Highlights:

• Top improvement: The _S_key function in the red-black tree implementation shows a 165.38% response time improvement
• STL optimizations: Major gains in C++ Standard Template Library functions, especially iterators and tree operations
• Consistent improvements: Most functions show both response time and throughput improvements
• Focus area: All changes are in build.bin.libllama.so, indicating targeted optimizations in the core library

The comparison between base version 04f5e641 and target version 8da5dfe1 reveals that this pull request successfully optimized memory management and data structure traversal patterns across the codebase.