About

Imported from SouthpawIN/turbofit (SKILL.md). Install upstream with npx skills add SouthpawIN/turbofit. Copyright stays with the author (MIT).

Turbofit

Name: turbofit
Author: SouthpawIN

2.3 model authority

Fit List keeps dedicated VRAM separate from integrated/RAM-only total memory. Dedicated: Maple Preview TQ2_0 at 8GB, Qwen 3.8 27B Unleashed UD-IQ3_XXS at 16GB, UD-Q3_K_XL at 24–95GB, and Qwen 3.8 27B 16-bit at 96GB+ until an Unleashed FP16 GGUF is published. Shared total memory: Maple at 8–15GB, Ornith 1.5 35A3B at 16–23GB, and Unleashed UD-Q3_K_XL at 24GB+. An 8GB GPU may also use Ornith when host RAM can hold offloaded experts. Maple remains Auto on dedicated 8GB at native 64K/128K. Auxiliary is Ornith, optional Carwin Nano, or auto.

Use this bundled plugin skill when configuring or inspecting Turbofit for Hermes Agent.

Operator workflow

Call turbofit_status to inspect provider registration, gateway health, selected hardware profile, active rung, and stable routes.
Call turbofit_configure with profile: auto for hardware selection. Manual hardware-*gb profiles are accepted only when physical topology fits.
Set primary: true to use custom:turbofit with model auto as the main Hermes provider.
Set fallback: true to append Turbofit to the canonical fallback_providers chain; set it false to remove only Turbofit while preserving other fallbacks.
Set publish_tailnet: true to create private Tailscale Serve routes for the provider and dashboard; the returned HTTPS provider URL is registered automatically.
Set install_sirvir: true to install or update the canonical SouthpawIN/sirvir GitHub profile without replacing its memories or user state.
Set install_freetoken: true only on Linux x86_64 + NVIDIA driver 580+ + CUDA toolkit 13+ to install pinned FreeToken 0.1.2 as a text-only MoE candidate. It never changes Auto until exact on-box campaigns promote a supported model recipe.
Start a new Hermes session after provider changes.

/turbofit setup refreshes Hermes Desktop. A refused http://127.0.0.1:8091/v1/models means the Turbofit runtime is down — not the Hermes messaging gateway. Diagnose that from Sirvir or Desktop with turbofit_status. Setup downloads recommended models if they are missing.

Intelligence benchmarks

scripts/turbofit-catalog-campaign proves native runtime fit and TPS; it does not produce intelligence scores.
scripts/turbofit-intelligence-campaign runs the exact successful quantized production recipe through pinned DeepSWE and the Turbofit agentic main/auxiliary pair harness.
Use status, run-one, or run --limit N; state is resumable in references/intelligence-campaign-state.json.
Scores require both benchmark suites and immutable raw evidence. Never replace missing scores with catalog tiers, parameter counts, or vendor benchmark claims.
/turbofit tiers and scripts/turbofit-hardware-tiers show every 8/16/24/48/64/96/200/300 GB class with pending versus measured intelligence and TPS.
scripts/turbofit-intelligence-campaign benchmarks only the current machine's TurboFit List tournament candidates. rebuild-scores recomputes derived composites from raw suite counts; zero-call/token trials remain invalid infrastructure.
scripts/turbofit-promote-list-winner promotes only an exact-tier candidate with current physical evidence, positive intelligence/TPS/balanced values, and matching recipe hashes. scripts/turbofit-list renders the global evidence-only List.
Qwen 3.8 DFlash2 is a separate candidate runtime/artifact pair (dflash2-llama.cpp, Qwen3.8-27B-DFlash2-Q4_K_M.gguf). Never attach that drafter to Bonsai. Bonsai uses its own released DSpark sidecar and Prism runtime until a dedicated Bonsai DFlash checkpoint exists.

Portable memory allocation

Hardware fingerprints classify memory as dedicated, unified, or cpu and reserve 5% of host RAM, bounded to 1–8 GiB.
Dedicated systems may combine accelerator VRAM with host RAM through llama.cpp offload; contexts beyond the model's native window place KV cache in host RAM when at least 32 GiB is usable.
Unified-memory systems count RAM once and suppress discrete multi-GPU split flags.
CPU-only systems set model and draft GPU layers to zero.
Native backend order is CUDA, ROCm, Vulkan, then CPU on Linux/Windows, and Metal on macOS. Use scripts/install-native-runtimes --backend <backend> for an explicit build.

Invariants

Stable model IDs are auto, active:main, and active:aux.
FreeToken support is candidate-only: no active Qwen 3.8/Ornith replacement, no source TPS inheritance, and no Auto promotion without exact hardware/model evidence.
External GPU processes are read-only pressure signals and are never terminated or signaled.
The hardware recommendation remains the healing ceiling; transient pressure changes only the effective rung.
Runtime activation and model lifecycle remain owned by NativeRuntimeBackend, which signals only PID-verified children it launched.
Plain HTTP provider endpoints are limited to loopback or Tailscale addresses; all other endpoints require HTTPS.
Every native llama.cpp command includes --jinja; DSpark variants include target, draft, projector, and draft-attention arguments.

Use it

Copy one of these into your project. Installing also returns the manifest and these snippets.

yaml

targets:
  - https://api.opensmartroute.ai/api/v1/registry/southpawin-turbofit-turbofit/manifest   # or paste the manifest below

Manifest

An Open Capability Manifest: the router reads it to know what this does, what it costs and when to pick it.

southpawin-turbofit-turbofit.ocm.jsonjson

{
  "ocm": "1",
  "id": "southpawin-turbofit-turbofit",
  "kind": "skill",
  "name": "turbofit",
  "description": "Operate the Turbofit adaptive local inference plugin.",
  "publisher": "SouthpawIN",
  "version": "1.0.0",
  "capabilities": {
    "domains": [
      "general"
    ],
    "tags": [
      "skill-md",
      "hermes-agent",
      "plugin",
      "llama-cpp",
      "adaptive-runtime",
      "github"
    ],
    "languages": [
      "en"
    ]
  },
  "quality_prior": 0.6,
  "examples": [
    "Operate the Turbofit adaptive local inference plugin."
  ],
  "primary": false,
  "metadata": {
    "source": {
      "provider": "github",
      "repository": "https://github.com/SouthpawIN/turbofit",
      "path": "SKILL.md",
      "ref": "98b45598785c4ca8efe5a5d5ea0835782f4ee007",
      "url": "https://github.com/SouthpawIN/turbofit/blob/98b45598785c4ca8efe5a5d5ea0835782f4ee007/SKILL.md",
      "key": "SouthpawIN/turbofit/SKILL.md"
    },
    "license": "MIT"
  },
  "instructions": "# Turbofit\n\n## 2.3 model authority\n\nFit List keeps dedicated VRAM separate from integrated/RAM-only total memory. Dedicated: Maple Preview TQ2_0 at 8GB, Qwen 3.8 27B Unleashed UD-IQ3_XXS at 16GB, UD-Q3_K_XL at 24–95GB, and Qwen 3.8 27B 16-bit at 96GB+ until an Unleashed FP16 GGUF is published. Shared total memory: Maple at 8–15GB, Ornith 1.5 35A3B at 16–23GB, and Unleashed UD-Q3_K_XL at 24GB+. An 8GB GPU may also use Ornith when host RAM can hold offloaded experts. Maple remains Auto on dedicated 8GB at native 64K/128K. Auxiliary is Ornith, optional Carwin Nano, or auto.\n\nUse this bundled plug",
  "cost": {
    "context_tokens": 1315
  }
}

Fetch it by URL: GET /api/v1/registry/southpawin-turbofit-turbofit/manifest?version=1.0.0

Reviews

Star ratings from people who tried it. One review per account; edit yours any time.

No reviews yet. Install it, try it, and be the first to rate it.