server : fix 501 on multimodal models blocking text-only slot save/restore (#21133)

[CHIPMUNK-T0T]

Overview

When I used a multimodal model (e.g. Qwen3.5) and load and use it with an image modality with mmproj, this llama.cpp server returned HTTP 501 unconditionally even for text-only conversations, because the check looked at model capability (mctx) rather than the slot's actual content.
As a result, the prefill cache could not be stored or reused even for text-only conversations, which leads to slow TTFT for long prompts.
The affected operations were /slots save / restore / erase on a server with --mmproj loaded. These now gate on the slot's content (has_media()): a text-only slot is allowed, and only a slot that holds media is rejected.

Additional information

This PR helps downstream consumers (e.g. Ollama) that reuse a prefilled cache across requests via slot save/restore on a multimodal model, which was previously blocked whenever an mmproj was loaded.
Only the text-only case is newly allowed; all other behavior is unchanged — slots that actually hold media are still rejected, and text-only (non-multimodal) servers are unaffected.

Notes:

• The restore path keeps using the existing llama_state_seq_* format. Since that format cannot store media chunks, a restored slot is always text-only, so no media gate is needed there.
• On save, get_text_tokens() is used instead of get_tokens() (which asserts !has_mtmd); the slot is confirmed media-free first.
• Serializing image/audio chunks is out of scope here and could be a follow-up.

Testing:

• New tests on a multimodal model (tinygemma3): text-only save / restore / erase succeed, while saving a slot that holds an image is rejected with HTTP 501. The existing text-only tests are unchanged.
• Local run: clean build (-j8), test_slot_save.py passing 5/5.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, but I used AI translater using Ai agent, because I'm Japanese, not a native English user, so it is difficult to tell you my soul!
  • The original motivation came from me. I used ollama to more fast decode/ttft, and I had AI aided (Codex and Claude Code) investigate and assess its impact/scope, but I made the final decision on every implementation change myself.
  • I have reviewed all changes, understand them fully, and can explain any line without AI assistance.

ISSUE #21133

[ggml-gh-bot]

Hi @CHIPMUNK-T0T, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[CHIPMUNK-T0T]

Thanks for the notice.

To clarify my AI usage: I am a native Japanese speaker, so it is difficult to tell you my native opinion, so I thought the original explanation and reasoning about this topic in Japanese, and I used AI assistance to translate and polish the English text.
The PR design, implementation decisions, final commit contents, and tests are mine.
I reviewed the final English description and code myself, and I can explain and maintain the changes.

If this still does not satisfy the project policy, I am happy to shorten or rewrite the PR description in simpler wording.

[CHIPMUNK-T0T]

Hi @CISC

this PR is still ready for review and currently mergeable.

The main use case is reusing text-only prefill cache when a multimodal model is loaded, while continuing to reject slots that actually contain media.

Please let me know if the scope or implementation should be adjusted.

[ggerganov]

Thanks, this is a good improvement.

We should also support save/restore with the media. This simply requires to serialize the server_tokens together with the llama state. In case you are interested for a follow-up PR.

[ggerganov]

Please rebase this branch to the latest master.