v1.7.0

What's new

Decision models. llama.cpp's /v1/systemone serves models that answer a question about an input with calibrated probabilities instead of text: a choice between options, a score on a scale, or a true/false probability. Quartermaster now runs them as their own model type.

• Discovered automatically. A GGUF carrying <arch>.decision.type (Kev, OpenJev, Laya, Clef, ...) gets its own config entry, with its context capped at 8192 and whole-batch encoders launched with -b/-ub at the full window.
• POST /v1/systemone is proxied through the scheduler and API keys like any other route, and llama-server validates the body. Decision models are listed on /v1/models with output_modalities: ["decisions"].
• Decision tab in the playground. Write a prompt, ask one question (Choice, Score or True / false) and read the answer as a probability bar per option or level. A projector next to the model enables image attachments.
• Decision tab in the model browser finds them on Hugging Face, and a wiki article explains the request and answer shapes.
• Outdated llama.cpp is flagged. A llama-server without the route (upstream before b11361) is marked in the dashboard instead of loading the model as a plain LLM that 404s every question.
• VRAM sized like a chat model. The estimate charges the compute buffer at the model's real batch size, and the full attention-score matrix for whole-batch encoders, instead of a flat pad.

After upgrading: the config is regenerated once (genVersion v98 to v100).

Changes

• autogen: bump genVersion to v100 for decision estimates
• autogen: size decision compute buffers by ubatch
• ui: drop the model path from the decision answer footer
• ui: polish the decision tab
• ui: decision tab as a prompt + one question form
• server: wiki article for decision models
• ui: fork icon for decision models
• hub: find decision models in the browser's Decision tab
• ui: Decision model category and playground tab
• server: flag decision models whose llama.cpp predates /v1/systemone
• server: advertise decision models on /v1/models
• autogen: discover llama.cpp decision models
• server: let llama-server validate /v1/systemone bodies
• server: pass-through /v1/systemone with Typesafe contract guard

Full changelog: v1.6.1...v1.7.0

Instructions

Download the setup program for your platform:

• Windows -- quartermaster-setup-windows-amd64-v1.7.0.exe
• Linux -- quartermaster-setup-linux-amd64-v1.7.0 (or -linux-arm64-v1.7.0)
• macOS, Apple silicon -- quartermaster-setup-darwin-arm64-v1.7.0

The unix ones need chmod +x first. There is no signed macOS build yet, so
clear the quarantine flag before the first launch:

xattr -d com.apple.quarantine ./quartermaster-setup-darwin-arm64-v1.7.0

SHA256SUMS covers every file here.

The other files, and who they are for

quartermaster-windows-amd64.exe and quartermaster-linux-amd64 /
-linux-arm64 / -darwin-arm64 are the bare server binary, not an
installer. Nothing about them creates a Start Menu entry, a desktop shortcut or
an uninstall record. They are published for two reasons:

• the in-app updater downloads one to replace the running binary, so a release
  without them silently stops every existing install from updating;
• a headless Linux or macOS box can run one directly: chmod +x, point it at
  a models folder, and install backends from Settings.

Installing for the first time? Take a quartermaster-setup-... file from
the instructions above.