mostlygeek/llama-swap v272 on GitHub

Three compatibility fixes and a couple of UI improvements.

This release is a collection of some UI improvements, compatibility
enhancements and a bit of work to make content in this repo easy to import and
reuse in the new https://llama-swap.ai website.

Thanks to everyone who filed detailed bugs and contributed changes. The
quality of this project would not be where it is without the community.

1249e7c changelog: make into package: makes changelog/ an importable Go package with Versions(), Latest() and Get() over the embedded per-release markdown so llama-swap.ai can display the release notes by @mostlygeek
PR #1315 ui: reduce ActivityTable render rate: throttles the Activity table to one render every 125 ms and skips redraws while the inflight panel is hidden, because live token/duration counters repainted on every update and burned CPU whenever several streams were active by @mostlygeek
PR #1318 UI: category and a capability badge for decision models: gives models declaring the decision capability their own sidebar/list category and a matching badge, since categories were derived from capabilities but decision models fell through with neither by @sousekd
PR #1319 internal/server: pass through 1xx interim responses in recorders: forwards 1xx statuses without recording them, because Expect: 100-continue (sent by curl for bodies over about 1 MiB) made the recorders latch onto the relayed 100 Continue and discard the real status, so upstream 400/500 errors arrived at the client as 200 by @mostlygeek
PR #1326 Remove metadata filtering to allow custom fields alongside capabilities: stops dropping architecture, capabilities, supported_parameters, context_length and context_window from user metadata when capabilities are set, since that metadata is nested under meta.llamaswap and cannot collide with the generated top-level fields by @mostlygeek
PR #1327 Add autoload=false support to /props endpoint: answers GET /props?autoload=false with 409 instead of loading a model that is not running (as llama-server's router mode does), so clients like llama.vscode that probe every listed model no longer load them one after another and evict the loaded model by @mostlygeek