PATCH_NOTES.md

Minimal patch notes

If the existing Leon4gr45/llama-swap Space already starts a recent llama-swap + llama-server successfully, the smallest change is:

1. Replace its current config.yaml with the supplied config.yaml.
2. Add the HF_TOKEN Space secret after accepting the Gemma license.
3. Rebuild/restart.
4. Test GET /v1/models, then POST /v1/chat/completions.

You do not need to replace the existing Dockerfile merely to change Bonsai/MiniCPM -> Gemma if its current llama.cpp build is recent enough.

Use the supplied Dockerfile only if you want the Space reduced to a clean, generic llama-swap CPU deployment based on the current upstream image.

Resource-first defaults

• model: google/gemma-3-4b-it-qat-q4_0-gguf:Q4_0
• exposed model id: gemma-3-4b
• context allocation: 8192
• parallel slots: 1
• CPU threads: 2
• KV cache: q8_0
• idle TTL: 300 s
• text-only by default (--no-mmproj)

Multimodal

Copy config.multimodal.yaml over config.yaml to allow llama.cpp to load Gemma's multimodal projector. This costs additional download time and memory, so it is intentionally not the default.