Skip to content

CHANGELOG.md

Latest commit

 

History

History
125 lines (110 loc) · 5.91 KB

CHANGELOG.md

File metadata and controls

125 lines (110 loc) · 5.91 KB
Raw

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Fixed

  • up/restart no longer pass --system-prompt to llama-server, which dropped the flag upstream (ggml-org/llama.cpp#9811) and rejects it with error: invalid argument: --system-prompt, failing every start when chat.system_prompt was set on a mode: server config. The flag is still sent for mode: interactive (llama-cli kept -sys/--system-prompt). A warning is now printed when chat.system_prompt is set on a server config, since it silently has no effect there — send it as a "system" role message in /v1/chat/completions requests instead.

[1.2.0] - 2026-05-04

Added

  • llmconfig update — self-update command. Downloads the latest release from GitHub, verifies its SHA256 against checksums.txt, and atomically replaces the running binary (the previous binary is kept as <binary>.old). Flags: --check (no install), --version <tag> (pin a specific release / downgrade), --force (reinstall on current). Updates the llmc alias binary in the same operation when present. Refuses to download from anywhere outside the official GitHub releases URL.
  • .llmc is now the canonical extension for config files. New templates, init, add, pull, and archive imports all write .llmc. Built-in templates were renamed (e.g. templates/gemma.yaml → templates/gemma.llmc); files are still YAML inside.
  • Auto-migration on first run: existing ~/.llmconfig/configs/*.yaml files are renamed to *.llmc with a *.yaml.bak backup left alongside, gated by a marker file so the scan only runs once. Conflicts (both extensions for the same name) are skipped with a warning.
  • internal/config exports ConfigPath, FindConfigInDir, ListConfigNames, ListConfigPaths, IsConfigFile, and TrimConfigExt so cmd/ doesn't have to repeat extension-aware globs.

Changed

  • Loader (internal/config.Load) now searches for both .llmc and .yaml, with .llmc winning when both exist. Bare-name lookups via CLI args resolve transparently to whichever file is present.
  • Archive bundles (.llmcpkg) now ship configs as configs/<name>.llmc. Reading still accepts legacy bundles whose inner config is .yaml; on import the file lands as .llmc on disk regardless.
  • llmconfig version --check now points users at llmconfig update instead of the bootstrap install scripts.

[1.1.0] - 2026-05-04

Added

  • ik_llama backend — drop-in alternative to llama that runs the ikawrakow/ik_llama.cpp fork (SOTA quants, MLA, fused MoE). Set backend: ik_llama in any config; ik-only flags live under a new optional ik_llama block (rtr, mla, fmoe, ser, cuda_graphs).
  • llmconfig install ik_llama — clones and cmake-builds ik_llama.cpp into ~/.llmconfig/cache/ and installs llama-server / llama-cli into ~/.llmconfig/bin/ik-llama/. Supports --backend cpu|cuda, --ref <tag|commit>, --jobs N, --verbose, and --file <archive> (bring-your-own-binary). Build log lands at ~/.llmconfig/logs/install-ik-llama.log.
  • llmconfig ik_llama — status command (path, version) mirroring llmconfig llama.

1.0.0 - 2026-05-03

First stable release. The 1.0 line stabilizes the CLI surface (commands, flags, config schema, on-disk layout) and ships with a documented contributor workflow. Future breaking changes will go through a deprecation cycle and a major version bump.

Added

  • gateway command — a unified OpenAI-compatible HTTP gateway that routes requests to the right running model based on the "model" field in the request body.
  • state prune command — marks running entries with dead PIDs as stopped in the state file.
  • Bare --template flag on init — running llmconfig init --template with no value opens an interactive template picker.
  • Templates: flux-schnell, flux-dev, phi4, plus seven additional unsloth-based templates tuned for 16 GB VRAM. 18 built-in templates total.
  • Embedded templates (templates/*.yaml via go:embed) — one source of truth for both the picker and --template=<name>.
  • Interactive pickers and shell autocomplete for every command.
  • LICENSE (MIT), CONTRIBUTING.md, CHANGELOG.md, and GitHub issue / pull-request templates.
  • docs/ folder with reference.md (full reference) and templates.md (template catalogue).

Changed

  • Cache directory and command renamed: the on-disk directory is now ~/.llmconfig/models/ and the management command is llmconfig files (previously llmconfig cache).
  • llmconfig rebranded from llamaconfig. The shorter llmc alias works everywhere too.
  • Documentation reorganized: README slimmed to fit a single screen, long-form docs moved into docs/. The reference list of templates was also corrected to match what actually ships.

Fixed

  • Downloader now writes to <file>.tmp and renames to the final filename on completion, so an interrupted download is never mistaken for a complete one.
  • File handle is closed on the early-return path in the downloader.
  • Gateway returns clearer errors when no models are running or when the requested model is not available.
  • ESC binding cancels all huh forms cleanly (pickers and confirm prompts).
  • Hardware-aware health check accepts 401 and 403 as "ready" for auth-protected backends.
  • HuggingFace 401/403 errors point at the exact repo page; falls back to the hf CLI's on-disk token when env vars are unset.
  • Concurrency and safety: state-file locking, signal handling, archive collision handling.