UPSTREAM PR #17765: Add a search field on model selector / improve mobile display

[loci-dev]

Mirrored from ggml-org/llama.cpp#17765

Make sure to read the contributing guidelines before submitting a PR

Add a search field to the model selector / Improve mobile display / Simplify the code

Close #17741

[loci-agentic-ai]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #438

Analysis Overview

This PR introduces frontend-only modifications to the llama-server web interface, specifically adding search functionality to the model selector dropdown and improving mobile display responsiveness. The changes affect 3 files: a Svelte component, a TypeScript constants file, and a compiled HTML artifact.

Performance Impact

No performance impact on inference operations. All modifications are isolated to the web UI layer with zero changes to core inference binaries or functions. The power consumption analysis confirms stable performance across all backend components:

• libllama.so: 194,027 nJ (0 nJ change)
• libggml-cpu.so: 117,027 nJ (0 nJ change)
• libggml-base.so: 59,071 nJ (0 nJ change)
• llama-run: 218,706 nJ (0 nJ change)

Tokens per second: Unaffected. No changes to llama_decode, llama_encode, or llama_tokenize functions. Response time and throughput for inference operations remain unchanged.

Binary-level changes: Two binaries show complete removal (llama-cvector-generator: -249,105 nJ, llama-tts: -253,822 nJ), but these are unrelated to this PR and represent separate build configuration changes in the version comparison.

The PR adds client-side search filtering with O(n) complexity per keystroke, which is negligible for typical model counts and occurs entirely in the browser without backend interaction.