feat(server meta): expose SWA geometry (n_swa + SWA-layer count) for exact gemma-4 KV prediction

[marksverdhei]

Extends #119 (which added n_layer/n_head_kv/n_embd_k_gqa/n_embd_v_gqa/cache_type_{k,v} to /v1/models meta). Those let a client compute KV/token assuming all layers scale with context — correct for dense models, but a 2–8× over-estimate for SWA-hybrid models (gemma-4 12B/31B, gemma-4-12b-256k@gem), where most layers cap KV at a small sliding window and only a few full-attention layers scale with ctx.

heierchat’s VRAM panel currently uses the all-layers value as a safe upper bound (errs toward over-VRAM, never under-predicts an OOM) labeled ≈. To make gemma-4 exact the FE needs the SWA split, which llama.cpp already knows.

Proposed additive meta fields (per loaded model, same path as #119)

field	source	example (gemma-4-12b)
n_swa	llama_model_n_swa(model) (already public)	1024
n_swa_layers	new accessor llama_model_n_swa_layers(model) = count of il where hparams.is_swa(il)	e.g. 40
swa_type	hparams.swa_type as string (none/standard/chunked)	standard

n_full_layers = n_layer − n_swa_layers is derived on the FE.

Exact KV/token the FE will then compute

kv_per_tok(ctx) = (k_gqa + v_gqa) * kv_type_bytes * (
                    n_full_layers
                  + n_swa_layers * min(ctx, n_swa) / ctx )

i.e. full layers scale with ctx; SWA layers saturate at min(ctx, n_swa).

Implementation sketch (our fork — clean, mirrors #119)

1. src/llama-model.cpp + include/llama.h: add LLAMA_API int32_t llama_model_n_swa_layers(const llama_model * model) next to the existing llama_model_n_swa — loop il in [0, n_layer) counting hparams.is_swa(il). (n_swa, is_swa, is_swa_any already exist in src/llama-hparams.h.)
2. tools/server/server-context.h: add int32_t model_n_swa; int32_t model_n_swa_layers; std::string swa_type; to server_context_meta.
3. tools/server/server-context.cpp: populate them where feat(server): expose KV-cache geometry + quant in /v1/models meta #119 populated n_layer etc.; emit n_swa, n_swa_layers, swa_type in the /v1/models meta JSON.

Downstream (tracked separately, snoop-kube’s lane)

• omniswap: fold the 3 fields into meta.llamaswap (mirror the dims/KV passthrough from omniswap [Frontend] Add lora-adapter toggler / selection list for selected model #13).
• Roll: titan-llm + centurion-llm (ht image), llm-pool home+gems, and llama-pascal@gem (coordinate with crystal-mom). LOADED-models-only (same cold-meta limitation as feat(server): expose KV-cache geometry + quant in /v1/models meta #119).

Impact

None until rolled; purely additive meta. Makes the gemma-4 VRAM verdict exact instead of a safe upper bound. Related: #119, #122. FE side queued by heierchat ("Exact KV for SWA-hybrid models in the VRAM panel").