server : fix swa-full logic

[ggerganov]

Overview

fix #21468
alt #21749

Simplify the logic by augmenting llama_model_n_swa with a server_context.n_swa member. When --swa-full is passed, we set n_swa = 0 to simulate a non-SWA model

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[shipped-it]

I'm a bit unsure about the approach. Why would the model be reported as non-SWA? It is still SWA, just with a full-size cache.

Also, you've made the suggestion to use --swa-full --no-mmproj. Correct me if I'm wrong, but in practice I've seen repetitions issues when using Gemma 4.

I will confirm shortly if this fixes the issue, or if I get the repetition problem.

[ggerganov]

Also, you've made the suggestion to use --swa-full --no-mmproj.

Cache reuse does not work with mmproj.

Correct me if I'm wrong, but in practice I've seen repetitions issues when using Gemma 4.

I am not following. This fixes the cache reuse logic - I am not aware of any repetitions.

[shipped-it]

I confirm that it is fixed with this PR or #21749 (tested on ROCm)

Without PR:
warm req: prompt_n=821, prompt_ms=982

With PR:
warm req: prompt_n=5, prompt_ms=71 so about 13x faster