Add support for the locateanything mmproj projector type used by NVIDIA EAGLE-architecture VLMs (e.g. nvidia/LocateAnything-3B, yuuko-eth/ LocateAnything-3B-GGUF)

[mohanga-qcom]

Background

NVIDIA's EAGLE architecture VLMs use a custom projector type locateanything in their mmproj GGUF files. This type is not in the supported projector list in
llama.cpp's clip.cpp at the version bundled with GenieX v0.3.17 (commit ae9291e).

The LLM backbone (qwen2, 3.4B) loads and runs correctly. Only the vision encoder (mmproj) is blocked.

Affected models:

• nvidia/LocateAnything-3B and its GGUF quantizations (yuuko-eth/LocateAnything-3B-GGUF)
• Any other EAGLE-architecture VLM using locateanything projector

———

Request

1. Update the bundled llama.cpp to a version that includes locateanything projector support (check upstream llama.cpp clip.cpp for the relevant commit)
2. Or add locateanything to the projector type registry in the bundled clip.cpp

———

Reference

• NVIDIA EAGLE architecture: https://github.com/NVlabs/Eagle
• Model card: https://huggingface.co/nvidia/LocateAnything-3B
• GGUF quantization: https://huggingface.co/yuuko-eth/LocateAnything-3B-GGUF
• mmproj GGUF metadata key: clip.projector_type = locateanything

Tried on IQ8 device HTP v75.

[mansiverma897993]

Alan Zhu (@alanzhuly) opened #1326 tested with the real model report in the PR. Mind approving CI and reviewing when you have a moment?

[mengshengwu]

Upstream track: ggml-org/llama.cpp#24749