v3.2.13

v3.2.13: mmap expert caching, launch hardening, safer downloads, Wind…

v3.2.12

This maintenance release hardens the install, recommendation, download and launch flow.

• Model downloads pin a repository revision and verify every file's size and content hash, including reused files. Native transfers stop after three minutes without progress and preserve partial data for retry.
• The installer retains the recommended quant. Catalog generation reads the current Artificial Analysis payload and preserves main-model parameter/MoE evidence while excluding draft heads and sparse indexers.
• Backend activation accepts current IK's normal help exit convention and verifies real GPU buffer enumeration when its CLI lacks --list-devices. Required flags, architecture and accelerator checks remain mandatory.
• Claude-mode placement keeps backend fitting within the allocation guard. Automatic context accounts for a resumed session's transcript.
• Linux launch can relaunch once after a learned verification OOM, after releasing the failed process. Failed or unrecorded recovery still stops clearly.
• IK logging no longer raises verbosity in a way that can expose request bodies.

Hot experts remain off by default. Windows GPU serving is still unverified; macOS is deprecated and best-effort. Model tool-use quality and CPU-only task time remain model dependent. This release does not claim maximum hardware utilization or universally optimal tuning.

v3.2.11

ggrun v3.2.11

v3.2.10

ggrun v3.2.10

v3.2.9

ggrun v3.2.9

Fixes the broken Linux and macOS bundles (#28), the Windows prebuilt CUDA
lookup (#27), and defects that stopped ggrun serving small or unusual
models.

v3.2.8

- Explicit 'off' reviewer option (--claude-reviewer off / TUI selector)
- Auto reviews fall back to the main model when no/overflowed reviewer
- Installer no longer aborts on fork-named llama-server (#26)
- Config-screen lag, self-review, resume-across-model, decode-in-status fixes
- CUDA backend: keep the bundle, use a backend that starts, fetch NCCL

v3.2.7

try 1/3: CUDA stub on default linker path

v3.1.0

ggrun v3.1.0

ggrun is now a single Go app you install in one command and run anywhere.

Install

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/raketenkater/ggrun/main/setup.sh | bash

# Windows (PowerShell)
iwr -useb https://raw.githubusercontent.com/raketenkater/ggrun/main/install.ps1 | iex

Prebuilt bundles (Linux CPU/Vulkan, macOS arm64 Metal, Windows x86_64 CPU) are
downloaded and verified against SHA256SUMS. Linux CUDA/ik_llama.cpp builds
from source for your GPU arch.

What's new since v3.0.0

• Exact-ledger multi-GPU MoE placement. Large MoE models now load reliably
  on heterogeneous multi-GPU rigs instead of over-committing the smallest card.
  The launcher emits --tensor-split and -ot from an exact per-GPU VRAM
  ledger — measured CUDA-context + compute-buffer overhead plus GGUF-exact
  non-expert, KV, and expert sizes.
• Interactive GUI overhaul. The whole TUI is navigable with arrow keys and
  Enter (letter hotkeys still work). New Settings screen lists every config
  option with its current value, backend selection is arrow-select, and the
  advanced launch screen is fully navigable.
• Reliable shutdown. Ctrl+C exits promptly; the keep-alive recovery loop
  no longer treats a requested shutdown as a crash.
• Config is the single source of truth. The installer no longer exports
  per-setting environment variables that silently shadowed the config file.
  Only LLM_APP_HOME and PATH are exported.
• One launcher binary. Installs a single ggrun command. ggrun
  with no arguments opens the GUI.
• Native Windows — PowerShell installer (install.ps1), Windows process/
  signal handling, CUDA backend via official llama.cpp prebuilts or local build.
• Apple Silicon Metal actually engages — Macs previously detected no GPU and
  ran CPU-only (-ngl 0); detection now synthesizes a unified-memory GPU and
  offloads to Metal. (Perf unvalidated on hardware — runs, not yet benchmarked.)
• Community tune pool — first launch reuses configs measured by others on the
  same GPU set, sanitized to safe performance flags. LLM_COMMUNITY_TUNES=off.
• Benchmarked vs llama.cpp --fit — on a 3090 Ti + 4070 + 3060 rig at 32k
  context, the fresh release matrix has 4B at 151.4 tok/s by default (185.7
  AI-tuned) vs 103.3, 27B at 37.4 by default (39.8 AI-tuned) vs 24.3,
  and Qwen3.5-122B-A10B at 23.1 AI-tuned vs 21.0. See
  docs/launch-performance.md for the current table and docs/performance.md for
  dated historical runs.

Known limitations

• macOS Metal performance not yet validated on real hardware.
• AMD: Vulkan path works; ROCm VRAM detection is still basic.
• Native Windows CUDA bundle is optional/local-build (no prebuilt CUDA asset).

Full changelog: CHANGELOG.md

v3.0.0

v3.0.0

v2.1.0 — keep-alive, backend selector, long-tune

New features since v2.0.0

• --keep-alive — never-give-up restart loop for unattended deployments. Also available as [k] toggle in llm-server-gui, or set LLM_KEEP_ALIVE=1.
• --backend llama|ik_llama — explicit backend selection, overrides auto-detection. llm-server-gui now shows an interactive backend picker and persists your choice to ~/.config/llm-server/config.sh.
• --tune-long-timeout — doubles per-round timeouts during --ai-tune, for very large / slow-loading models.
• AI Tune "baseline wins" result now persists to cache — future launches skip re-tuning instead of retrying.
• Interactive rounds prompt in gui after [a] AI Tune toggle (1-20, default 8).

Fixes

• --kv-quality no longer errors under set -u when the value is missing.
• Removed duplicate update check in gui (single check at startup).