UPSTREAM PR #18169: presets: refactor, allow cascade presets from different sources

[loci-dev]

Mirrored from ggml-org/llama.cpp#18169

Alternative to ggml-org/llama.cpp#17959

Fix ggml-org/llama.cpp#17948

Before this PR, the logic for loading models from different sources (cache / local / custom ini) was quite messy and doesn't allow ini preset to take precedence over other sources.

With this PR, we unify the method for loading server models and presets:

• preset.cpp is responsible for collecting all model sources (cache / local) and generate a base preset for each of the known GGUF
• preset.cpp then load INI and parse the global section ([*])
• it is then up to downstream code (e.g. server-models.cpp) to decide how to cascade these presets

The current cascading rule can be found in server's docs:

1. Command-line arguments passed to llama-server (highest priority)
2. Model-specific options defined in the preset file (e.g. [ggml-org/MY-MODEL...])
3. Global options defined in the preset file ([*])

[loci-review]

Explore the complete analysis inside the Version Insights

Pull Request #614 Performance Analysis Summary

Overview

PR #614 implements a preset configuration system refactoring for the llama-server multi-model router. The changes introduce a unified preset loading mechanism with cascading configuration support, affecting 6 files with 347 additions and 260 deletions. Performance analysis reveals no material impact on inference operations.

Key Findings

Performance Impact on Inference:
No inference-related functions were modified. The functions showing performance variations (std::vector iterators, std::chrono::duration_cast) are utility operations in configuration parsing code, not in the inference pipeline. Functions critical to tokens-per-second performance (llama_decode, llama_encode, llama_tokenize) remain unchanged with 0 ns delta in response time and throughput.

Affected Components:

• Configuration Management: New common_preset_context class centralizes preset loading from cache, local directories, and INI files
• Server Model Loading: Refactored server_models::load_models() implements three-tier cascading: global preset → model-specific preset → CLI arguments
• Argument Parsing: Added boolean argument handling in common_params_to_map() for negative flags

STL Iterator Performance Variations:
Analysis identified 145-226% response time increases in std::vector iterator functions within llama-cvector-generator (199 ns vs 81 ns baseline) and corresponding improvements in llama-tts (84 ns vs 199 ns baseline). These variations stem from compiler inlining decisions in non-inference binaries. Absolute impact: 118 ns per call in configuration parsing code executed once at startup.

Power Consumption:
Binary-level analysis shows negligible changes: llama-cvector-generator (-0.03%), llama-tts (+0.028%). All other binaries including libllama.so (0.0% change) show no measurable power consumption variation.

Code Changes Analysis:
The refactoring introduces structured preset management without modifying inference kernels. New functions (load_from_cache(), load_from_models_dir(), cascade()) operate during server initialization, not during token generation. The unset_reserved_args() helper ensures router-controlled parameters are properly isolated from model-specific configurations.

Conclusion:
This refactoring improves configuration management architecture without affecting inference performance. The observed STL iterator variations are compiler artifacts in startup code paths, not runtime inference operations. Tokens-per-second performance remains unchanged as no tokenization or decoding functions were modified.

[loci-review]

Explore the complete analysis inside the Version Insights