llama.cpp Now Lets You Pick the Vision GPU

llama.cpp release b10541 added `--mmproj-device`, plus the short `-mmdev` form, so builders can select the backend device used by mmproj. The release also preserves the existing `MTMD_BACKEND_DEVICE` environment variable. The merged pull request reports the argument working across `llama-mtmd-cli`, `llama-cli`, and `llama-server`, and reports `preset.ini` support for named presets, the global `*` preset, and runtime switching. Why does that matter? The pull-request author describes using different placements for large and smaller model configurations: mmproj can move to an integrated GPU for one preset and remain on a discrete GPU for another. A commenter describes a dual-card setup using an RTX 4070 and GTX 1080 Ti, with two configurations available through one `llama-server` rather than two separate instances. The reality check is important. The packet reports only one mmproj device is allowed, and documents an RPC0 failure for the environment-variable path because the RPC backend is not loaded when that value is processed. It also provides no comparative latency, throughput, memory, or quality benchmark. This episode therefore treats the change as a placement and configuration improvement—not a guaranteed performance upgrade. Verification should follow the pull request’s own test logic: read the CLIP backend log to confirm the selected backend, then check that model vision works as expected. Measure performance on your own configuration only after those functional checks pass. Verified sources: llama.cpp release b10541: https://github.com/ggml-org/llama.cpp... Merged pull request #23255: https://github.com/ggml-org/llama.cpp... Sources used: https://github.com/ggml-org/llama.cpp... https://github.com/ggml-org/llama.cpp... https://github.com/ggml-org/llama.cpp... https://github.com/ggml-org/llama.cpp... Subscribe to CodeMotion for visual explainers about local AI, AI agents, model architecture, developer tools, and the systems behind modern AI.