Feat - add kimi 2.5 Vision

[SamuelOliveirads]

Ports Kimi K2.5 Vision support based on the request in issue #1264.

This implementation references logic from upstream PR 19170, PR 16891 and PR 17022, and includes small updates in clip.cpp.

Tested in the default webui using IQ1-KT with the f16 mmproj.

[ikawrakow]

Is the code for the newest neighbor and bilinear interpolation taken from llama.cpp?

[ikawrakow]

What was the purpose of this change? In the original it is clear that we need to have either a merged qkv_w, or each of k_w, q_w, v_w. In the new version it is perfectly fine to not have any of these, or to have a subset of the required tensors.

[SamuelOliveirads]

It is just a small refactor to align with current upstream llama.cpp. There, they try to load every tensor in clip_model_loader regardless of whether it exists or not, and moved the validation logic inside build_vit. The same applies to k_b.
This isn't strictly for Kimi vision, but rather to keep ik_llama updated with upstream and simplify future model ports. However, since this refactor could impact other models I haven't tested, I can simply revert this change if you prefer.

[ikawrakow]

Same comment as above

[ikawrakow]

It says nearest neighbor or bilinear. The code does neither of these two things, not to mention that it somehow assumes nearest neighbor interpolation is the same thing as bilinear interpolation.

[SamuelOliveirads]

In llama.cpp they have the full bilinear code, but it wasn't necessary for Kimi, so I didn't include it. Since the current codebase only uses nearest neighbor, I assumed existing vision models were already relying on that logic. However, since I modified how clip.cpp handles ggml_scale_mode, I can include the proper bilinear implementation as a safety measure to avoid breaking other vision models.

[SamuelOliveirads]

Is the code for the newest neighbor and bilinear interpolation taken from llama.cpp?

Yes, although I didn't port the bilinear implementation since Kimi doesn't use it.

[vaulter]

Is the code for the newest neighbor and bilinear interpolation taken from llama.cpp?

Yes, although I didn't port the bilinear implementation since Kimi doesn't use it.

Just wanted to highlight - qwen 3.5 also has vision modality, i'm not aware which implementation it uses - but it might make sense to check since for possble future implementation

[SamuelOliveirads]

Just wanted to highlight - qwen 3.5 also has vision modality, i'm not aware which implementation it uses - but it might make sense to check since for possble future implementation

Qwen 3.5 use the same vision logic from 3V with we already support

[Skelectric]

Hi I tested this with both qwen3-vl and kimi-k2.5 and can confirm it works. thanks for the PR!