Skip to content

Releases: jlbyh2o/paddock

v0.3.3

Choose a tag to compare

@github-actions github-actions released this 15 Sep 15:48

Full Changelog: v0.3.2...v0.3.3

v0.3.2

Choose a tag to compare

@github-actions github-actions released this 15 Sep 03:46

Full Changelog: v0.3.1...v0.3.2

v0.3.1

Choose a tag to compare

@github-actions github-actions released this 15 Sep 02:21

Full Changelog: v0.3.0...v0.3.1

v0.3.0

Choose a tag to compare

@github-actions github-actions released this 14 Sep 22:16

The headline is a second front end.

ft-man web

ft-man web serves the same App the TUI runs to a browser — the same telemetry poller,
engine supervisor, job runner and download tracker, with every screen, action and
confirmation identical on both. It is a long-lived daemon, so the natural pattern is to
drive a headless GPU box from a browser, start a two-hour conversion, and check on it from
a terminal later. Both can run at once and drive the same engine.

It ships with a systemd unit (contrib/ft-man-web.service) and defaults to 0.0.0.0:7979.
Note that, like ft serve itself, it has no authentication unless you set one — bind to
loopback or set [web] token on anything reachable from a network you don't trust.

Your library is the Hugging Face cache now

ft-man reads models--org--name/snapshots/<sha>/ out of $HF_HOME/hub, so a checkpoint
pulled by hf download, from_pretrained, Unsloth or another engine is already in the
list — under its repo id, at its real size. Downloads go back to the same place, delegated
to hf rather than reimplemented.

A quantized GGUF repo is presented as the builds it holds rather than as sixty files: the
Hub tab asks UD-IQ3_XXS or Q8_0 and selects the matching shards, the shared tokenizer
and the right projector.

Per-checkpoint serving defaults

Two things FreeToken has no flag for, both written into the checkpoint and both reversible:

  • Sampling defaultstemperature, top_p, top_k, merged into
    generation_config.json (never over it — that file also carries the stop token ids).
  • Chat templates — fetched from any Hugging Face repo, render-checked against the
    model's real tokenizer before being applied.

Dashboard

FreeToken checkout status and version, a running-version line, an update action, and a
summary of what upstream commits change — asked of the model you already have loaded.

Also

  • Multimodal ft serve flags for the families that grew image input.
  • Current FreeToken flag spellings emitted, with profiles saved against the old names
    migrated when they load.
  • Cache limits read under the keys and units the engine actually publishes.
  • A closed port is no longer reported as a fault while the engine is stopped.

v0.2.0 and v0.2.1 were released as commits but never tagged here, so this release spans
them.

Install

Linux x86_64, glibc 2.34 or newer (RHEL 9, Ubuntu 22.04, Debian 12 and later). No toolchain
needed — the browser UI is built into the binary.

tar xzf ft-man-0.3.0-x86_64-unknown-linux-gnu.tar.gz
install -Dm755 ft-man-*/ft-man ~/.local/bin/ft-man
ft-man --doctor

Full Changelog: v0.1.0...v0.3.0

v0.1.0

Choose a tag to compare

@github-actions github-actions released this 07 Sep 00:25