inferhost 0.7.2

Run any Hugging Face GGUF model on your own GPU — chat/vision, text-to-speech (OuteTTS), and image generation (SDXL/Flux/Z-Image/Qwen-Image via stable-diffusion.cpp), all behind one OpenAI-compatible endpoint. Pulls official upstream binaries, no compiling. Type `inferhost` and you're done.

inferhost

Turn any GPU box into a private, multi-modal AI server in one command. Chat, vision, speech, and image generation — Qwen, Llama, Flux, SDXL, OuteTTS and more — all behind a single OpenAI-compatible endpoint. inferhost pulls official upstream binaries (nothing to compile), auto-fetches the right model files for you, and hot-swaps models in and out of VRAM on demand, so one card serves a big LLM and image generation without you juggling a thing. You only ever touch a .env file and one keyboard-driven dashboard.

📖 Full documentation: https://amirrouh.github.io/inferhost/

Run any Hugging Face GGUF model on your own machine — text/chat, vision, embeddings, text-to-speech, and image generation — behind one OpenAI-compatible endpoint. inferhost is a small Python framework that wraps llama.cpp, stable-diffusion.cpp, and llama-swap behind a single LiteLLM gateway at http://<host>:9001/v1.

One binary, two modes:

Command	What it does
inferhost	Interactive TUI dashboard — add models, pin, watch logs.
inferhost start | stop | restart | status	Headless control of the same daemons. No terminal required.

Key features:

• Multi-modal, one endpoint: chat/vision (/v1/chat/completions), text-to-speech (/v1/audio/speech, OuteTTS), and image generation (/v1/images/generations, SDXL / Flux / Z-Image / Qwen-Image) — all on :9001, all OpenAI-compatible. The extra engines (llama-tts, sd-server) are fetched automatically when you add a model that needs them.
• Single endpoint, always on: LiteLLM is bundled (no extra required) and auto-starts on :9001.
• KV cache compression on by default: K=q8_0, V=q8_0 — ~2× compression of the f16 baseline with near-lossless quality. Override per axis from the TUI Settings screen or .env.
• Pin = load now, with VRAM guard: Pressing P immediately loads the model into VRAM. If it won't fit, inferhost warns you and asks you to unpin something else first.
• Prebuilt binaries from upstream, nothing to compile: inferhost pulls llama-server straight from ggml-org/llama.cpp releases (Vulkan / ROCm / SYCL / OpenVINO / CPU on Linux, Metal on macOS arm64). Pin the version or change the backend from the TUI.

uv tool install inferhost
inferhost                  # TUI: add a model, press 's' to start
# … or, once a model is registered, run everything headlessly:
inferhost start            # spawn llama-swap + LiteLLM as background daemons
inferhost status
inferhost stop

That's it. The first launch downloads the runtime binaries (llama-server + llama-swap) for you with a progress bar; then the dashboard opens and you can add, start, stop, and inspect models from the keyboard. On unattended hosts, the same inferhost start/stop/... subcommands run everything without a terminal.

What it does

• One-key serving of any GGUF model published on Hugging Face.
• Automatic quantization selection based on available VRAM (Q6 → Q5 → Q4 → IQ4 fallback).
• OpenAI-compatible API out of the box, including tool calling and vision for any GGUF that ships an mmproj-*.gguf (auto-downloaded alongside the main file).
• Text-to-speech via llama.cpp's llama-tts — add an OuteTTS model and get /v1/audio/speech (WAV) on the same gateway. See Text-to-speech.
• Image generation via stable-diffusion.cpp's sd-server — SDXL/SD1.5 (single-file), and Flux / SD3 / Z-Image / Qwen-Image (multi-file, assembled with a paste-repo→pick-from-list component picker). Runs under llama-swap so it swaps VRAM with your LLMs. See Image generation.
• Stacked speculative decoding for MTP-capable models — combines llama.cpp's --spec-type draft-mtp with --spec-type ngram-mod so MTP handles novel tokens while ngram-mod dominates on repeated patterns (code, function names, etc.). The MTP draft depth is tunable per model in Configure (--spec-draft-n-max).
• Honest context windows — the served and advertised window is read against the GGUF's native trained context on disk, so what agents see over the API (and in the Hermes cache) always matches what llama-server actually loaded.
• Multi-model support via llama-swap, which lazy-loads model backends on demand.
• Auto-detected hardware: NVIDIA CUDA, CPU, or Apple Silicon Metal (prebuilt assets); for Vulkan/ROCm, point INFERHOST_LLAMA_SERVER_PATH at your own binary.
• Live download progress for both runtime binaries and Hugging Face model files.
• One binary, two modes — inferhost opens the TUI; inferhost start/stop/restart/status controls the same daemons headlessly. Drop-in for servers, cron jobs, and anything without a TTY.
• Full control from the TUI — change ports, edit context size and GPU layers, watch status of every daemon. No editor, no YAML, no extra commands.
• All defaults still overridable through environment variables or a .env file — the TUI just writes another .env file at ~/.config/inferhost/inferhost.env so your changes survive restarts.

Installation

Requirements: Python 3.11+, Linux or macOS. NVIDIA CUDA, Linux CPU, or Apple Silicon Metal are the supported prebuilt targets.

inferhost is a CLI app, not a library — install it globally with uv tool (or pipx), not into a project's dependencies.

# Recommended — global, isolated, on your PATH
uv tool install inferhost

# Alternatives
pipx install inferhost
pip install inferhost            # only inside an existing venv

Note: In v0.4 and earlier, LiteLLM was an optional [gateway] extra (inferhost[gateway]). From v0.5 it is bundled — a plain uv tool install inferhost is all you need. The [gateway] extra still exists as an empty alias for one release to avoid breaking existing install scripts.

⚠️ Don't use uv add inferhost. uv add pins it as a project dependency, so you can only run it via uv run inferhost inside that one project directory. Use uv tool install so inferhost is a normal command on your PATH.

Upgrade

uv tool upgrade inferhost                # if installed with `uv tool`
pipx upgrade inferhost                   # if installed with pipx
pip install -U inferhost                 # if installed with pip

To pin a specific version:

uv tool install --force 'inferhost==0.5.0'

Uninstall

uv tool uninstall inferhost              # if installed with `uv tool`
pipx uninstall inferhost                 # if installed with pipx
pip uninstall inferhost                  # if installed with pip

Inferhost stores runtime binaries, logs, and the model registry outside the Python install. To wipe everything (binaries, llama-server logs, model registry — but not downloaded GGUFs, which live in the Hugging Face cache):

rm -rf ~/.local/share/inferhost ~/.config/inferhost

To also drop downloaded models: rm -rf ~/.cache/huggingface/hub/models--*.

Usage

There is exactly one command:

inferhost

This opens the TUI. On first launch it downloads llama-server and llama-swap with a progress bar. Afterward you land on the dashboard.

Keys

Key	Action
a	Add a Hugging Face model (downloads the GGUF + any mmproj-*.gguf for vision)
n	Rename the highlighted model's public alias (regenerates llama-swap + LiteLLM configs)
c	Configure the highlighted model: per-model -c (context), -ctk / -ctv (KV cache K/V quant), -ngl (GPU layers), --parallel, -fa, reasoning, reasoning budget, raw extra args, MTP draft tokens (--spec-draft-n-max), and pin. The dialog shows the model's native trained context read from the GGUF on disk; a -c above it is clamped on load and the advertised window matches what's actually served. Blank fields inherit the global Settings value.
P	Toggle pin on the highlighted model — pins load the model into VRAM immediately; unpinning unloads it. inferhost checks VRAM first and warns if it won't fit.
d / Delete	Remove the highlighted model from the registry
s	Start llama-swap
x	Stop llama-swap
r	Restart llama-swap
p	Open the Settings panel (ports, context, GPU layers, flash attention)
R	Refresh
q	Quit

The top of the dashboard shows two live status rows: a GPU bar (per-card VRAM bar, used / total, utilization — via nvidia-smi, hidden on non-NVIDIA boxes) and a status bar with the daemon dots, ports, the selected model's active ctx, and which model llama-swap currently has resident in VRAM (loaded: <name>). Both refresh every two seconds.

Adding a model

Press a, type a Hugging Face repo id (e.g. Qwen/Qwen2.5-7B-Instruct-GGUF), and press Enter. The TUI lists the available GGUF files, marks the recommended quant for your hardware, and shows a live progress bar while it downloads. The model is registered against llama-swap and ready to serve.

Pinning models (load into VRAM immediately)

Press P on a highlighted model to pin it. Pinning loads the model into VRAM right away — it does not wait for the next request. inferhost checks available VRAM before pinning; if the model won't fit, a modal appears asking you to unpin something else first. Press P again on a pinned model to unpin and unload it. The sidebar marks pinned models with a ★.

Renaming a model

The name shown in the model list is also the value clients send as the OpenAI model field. Press n to change it. inferhost rewrites the llama-swap and LiteLLM configs in one shot and, if llama-swap is running, restarts it so the new alias is reachable immediately. No need to edit any YAML by hand.

Changing ports and other settings

Press p to open the Settings panel. You can edit swap_port, gateway_port, default_ctx, gpu_layers, flash_attention, parallel_slots, reasoning, reasoning_budget, the KV cache quants (kv_quant_k / kv_quant_v — accepts any f16/q8_0/q5_*/q4_*/iq4_nl/off), and the llama.cpp version + backend (llamacpp_version, llamacpp_backend) directly. Saving writes a managed env file at ~/.config/inferhost/inferhost.env, so your changes persist across restarts. Press r afterwards to restart llama-swap with the new values.

Endpoint

The single OpenAI-compatible endpoint is the LiteLLM gateway on port 9001:

curl http://localhost:9001/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2.5-7b-instruct-q4-k-m",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Use the model name column from the dashboard as the model field.

Text-to-speech (/v1/audio/speech)

If you add a Hugging Face repo whose GGUF ships a WavTokenizer / vocoder alongside the model (e.g. an OuteTTS repo), inferhost auto-detects it, downloads it, and serves the model as a text-to-speech model on the same gateway:

curl http://localhost:9001/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "outetts-0.2-500m-q4-k-m", "input": "Hello from inferhost.", "voice": "default"}' \
  --output speech.wav

The OpenAI Python SDK works unchanged (client.audio.speech.create(model=..., input=..., voice="default")). Notes:

• voice is required when calling through the gateway (the OpenAI spec and LiteLLM both mandate it). Any value works — see below; the SDK always sends it.

• TTS runs via llama.cpp's standalone llama-tts binary (the only way to synthesize OuteTTS), bundled automatically on install/update. It has no resident-server mode, so each request reloads the model (a few seconds of overhead) — fine for occasional/scripted use.

• Output is WAV. The voice value is ignored unless it points at a llama-tts speaker file — the model's built-in speaker is used otherwise — but the field must still be present (see above).

• TTS models are marked ♪ [tts] in the dashboard. They don't run under llama-swap and can't be pinned — they're served on demand by a small inferhost-tts daemon that inferhost start brings up automatically whenever a TTS model is registered.

• Auto-detect only: the vocoder must live in the same repo as the model. Repos that publish the vocoder separately aren't picked up. If you previously added an OuteTTS model as a plain chat model, remove and re-add it so the vocoder is detected and the model is reclassified as TTS.

Image generation (/v1/images/generations)

inferhost bundles stable-diffusion.cpp's sd-server, so adding an image model lights up an OpenAI-compatible image endpoint on the same gateway. Add a model with the Image generation kind selected in the add screen (same "paste repo → pick from a list" flow as LLMs):

curl http://localhost:9001/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{"model": "stable-diffusion-v1-5-q4-0", "prompt": "a red apple on a table", "size": "512x512"}' \
  | jq -r '.data[0].b64_json' | base64 -d > out.png

Supported image models (via the bundled sd-server):

Family	Files	Verified on GPU
SD 1.5 / SDXL (incl. Turbo)	single-file	✅ SDXL-Turbo, 512² in ~2 s
Flux.1 (schnell/dev)	diffusion + VAE + CLIP-L + T5XXL	✅ Flux-schnell, ~4 s
Flux.2 Klein (incl. Bonsai-Image)	diffusion + VAE + Qwen3-4B (--llm)	✅ Bonsai 1-bit, ~2 s
Z-Image (Turbo)	diffusion + VAE + Qwen3-4B (--llm)	✅ ~11 s
Qwen-Image / Qwen-Image-Edit	diffusion + VAE + Qwen2.5-VL (--llm) + mmproj (--llm_vision)	✅ edit via /v1/images/edits

For the multi-file families, inferhost auto-fetches the right companions (see recipes below).

The OpenAI Python SDK works unchanged (client.images.generate(model=..., prompt=..., size=...)). Notes:

• Single-file (SD1.5 / SDXL): paste the repo, pick one .gguf/.safetensors → done. Multi-file (Flux / SD3): inferhost auto-detects and downloads the companion VAE / CLIP / T5 files shipped in the same repo. The sd-server binary itself is fetched automatically the first time you add an image model.

• Shares VRAM with your LLMs: image models run through llama-swap in the swappable group, so an image model and a large LLM evict each other — only one big model is resident at a time. The model lazy-loads on the first request.

• Parameters: resolution via size per request; steps/cfg/sampler are set as per-model defaults (the model's extra_args in Configure) or per request via sd-server's <sd_cpp_extra_args>{...}</sd_cpp_extra_args> block in the prompt.

• Quality: same model weights as ComfyUI → comparable txt2img, but a subset of ComfyUI's ecosystem (no full node graphs) and slower on Vulkan than CUDA-PyTorch.

• Multi-file models just work via built-in recipes. When you add a model from a known family — Flux.1 / Flux.2 Klein / Z-Image / Qwen-Image (and anything built on them, e.g. Bonsai-Image) — inferhost recognizes it and auto-downloads the correct VAE + text encoder(s) from known-good, non-gated repos, and sets sane sampling defaults (--steps/--cfg-scale). No need to know which files go where. Supported encoders: CLIP+T5 (Flux.1), Qwen/LLM text encoder (--llm, for Z-Image / Flux.2 / Qwen-Image) and the vision ViT (--llm_vision, for Qwen-Image-Edit).

• Anything unmatched: the component editor. For a model with no recipe, add the diffusion file then open Configure — image models get a slot editor (VAE, Text encoder, Vision encoder, CLIP-L/G, T5XXL) where each is filled with the same paste repo → pick from list flow; inferhost downloads and wires it.

  Example — Z-Image-Turbo (3 files from 3 repos): diffusion leejet/Z-Image-Turbo-GGUF, VAE second-state/FLUX.1-schnell-GGUF/ae.safetensors (a non-gated mirror — the official black-forest-labs/FLUX.1-schnell is gated), text encoder unsloth/Qwen3-4B-Instruct-2507-GGUF. Set --steps 8 --cfg-scale 1 in the model's extra args (it's a turbo/few-step model). Verified end-to-end.

• Image editing (Qwen-Image-Edit, Flux Kontext): the OpenAI edit endpoint is multipart/form-data, which the LiteLLM gateway doesn't route by model. Reach it directly on llama-swap, which forwards any path for a model: POST http://<host>:9090/upstream/<model>/v1/images/edits (image + prompt as form fields). txt2img still goes through the normal :9001 gateway.

Configuration

Every setting is overridable through environment variables or a .env file in the working directory. Copy .env.example for the full list.

Variable	Default	Purpose
INFERHOST_SWAP_PORT	9090	llama-swap listen port. Defaults to 0.0.0.0 so it's reachable from your LAN / Tailscale — override INFERHOST_SWAP_HOST=127.0.0.1 to keep loopback-only.
INFERHOST_GATEWAY_PORT	9001	LiteLLM gateway port — the user-facing OpenAI endpoint.
INFERHOST_TTS_PORT	9092	Port for the inferhost-tts daemon (serves /v1/audio/speech for TTS models). Only runs when a TTS model is registered. Defaults to 0.0.0.0; override INFERHOST_TTS_HOST=127.0.0.1 for loopback-only.
INFERHOST_SDCPP_VERSION	latest	Pin a stable-diffusion.cpp release tag (image generation). latest pulls the newest; the sd-server binary is fetched automatically when you add your first image model.
INFERHOST_SD_STEPS	0	Default diffusion steps baked into the image model's launch (0 = sd-server's own default). Override per-model via extra_args.
INFERHOST_SD_CFG_SCALE	0	Default CFG scale for image models (0 = sd-server default).
INFERHOST_SD_SAMPLER	(default)	Default sampler for image models (e.g. euler, dpm++2m); blank = sd-server default.
INFERHOST_KV_QUANT_K	q8_0	K cache type (-ctk). q8_0 is ~2× compression, near-lossless; f16 is lossless.
INFERHOST_KV_QUANT_V	q8_0	V cache type (-ctv). Same accepted values as K. Drop to q5_0 / q4_0 to save more VRAM.
INFERHOST_LLAMA_SERVER_PATH	(auto)	Path to a custom llama-server binary (e.g. a self-built CUDA binary).
INFERHOST_DATA_DIR	~/.local/share/inferhost	Binaries, logs, and PID files.
INFERHOST_CONFIG_DIR	~/.config/inferhost	Model registry and generated YAML.
INFERHOST_HF_CACHE	~/.cache/huggingface	Hugging Face model cache.
INFERHOST_GPU_LAYERS	99	-ngl value passed to llama-server.
INFERHOST_DEFAULT_CTX	8192	Default context length for new models (clamped to the GGUF's native trained context if smaller).
INFERHOST_MAX_OUTPUT_TOKENS	0	Completion cap advertised to agents as max_output_tokens. 0 advertises the full served window (llama.cpp shares one budget for input + output); set a positive N to cap it for frameworks that reserve output room.
INFERHOST_FLASH_ATTENTION	on	-fa flag for llama-server.
INFERHOST_PARALLEL_SLOTS	1	--parallel flag — concurrent request slots per llama-server instance. 1 = serial.
INFERHOST_REASONING	auto	--reasoning flag — thinking mode for capable models. on, off, or auto.
INFERHOST_REASONING_BUDGET	-1	--reasoning-budget flag — token cap on thinking. -1 = unlimited, 0 = none.
INFERHOST_LLAMACPP_BACKEND	auto	Force the prebuilt variant: vulkan, rocm, sycl, openvino, cpu, or metal. Upstream has no Linux CUDA build — pick vulkan on NVIDIA Linux.
INFERHOST_LLAMACPP_VERSION	latest	Pin a specific upstream llama.cpp release tag (e.g. b9320 or 9320).
INFERHOST_LLAMASWAP_VERSION	latest	Pin a specific llama-swap release tag.
INFERHOST_SPEC_DRAFT_N_MAX	2	MTP draft tokens per step (only used on MTP-capable models). Set to 0 to disable the MTP lane.
INFERHOST_SPEC_NGRAM_MOD_N_MATCH	24	Minimum matching sequence length before ngram-mod drafts.
INFERHOST_SPEC_NGRAM_MOD_N_MIN	48	Minimum context window ngram-mod searches back through.
INFERHOST_SPEC_NGRAM_MOD_N_MAX	64	Max draft tokens ngram-mod proposes on a strong match.

Architecture

   Client                inferhost
   ------                ---------
   Your app  --HTTP-->   LiteLLM gateway        llama-swap (loopback)   llama-server
                         :9001 (public)  --->   127.0.0.1:9090   --->   (llama.cpp)

• llama.cpp runs the inference via the official upstream llama-server from ggml-org/llama.cpp — the backend (Vulkan, ROCm, SYCL, OpenVINO, CPU, Metal) is picked by hardware probe and can be overridden in the TUI.
• llama-swap sits in front of multiple llama-server instances and lazy-loads them on demand. It binds loopback (127.0.0.1) only.
• LiteLLM is the single user-facing gateway — always on, always bundled, serving :9001.
• inferhost-tts (optional, only with a TTS model registered) wraps llama.cpp's llama-tts behind /v1/audio/speech; LiteLLM routes the gateway's speech requests to it.
• stable-diffusion.cpp sd-server (optional, only with an image model registered) runs under llama-swap — so image models lazy-load and swap VRAM with LLMs — and serves /v1/images/generations. No extra daemon; LiteLLM routes image requests through llama-swap to it.

Troubleshooting: Both endpoints are reachable on all interfaces by default (0.0.0.0). If you set INFERHOST_SWAP_HOST=127.0.0.1 and then curl http://<lan-ip>:9090/... fails, that override is the reason — switch back to 0.0.0.0 or use :9001 (the LiteLLM gateway).

Mouse clicks not working? If you run inferhost inside tmux, tmux must have mouse mode on or it eats the click events before the TUI ever sees them. Fix once with: tmux set -g mouse on (or add set -g mouse on to ~/.tmux.conf). inferhost detects this on startup and surfaces a warning toast. To disable mouse capture entirely, set INFERHOST_MOUSE=off in your .env.

Development

The repo ships a run.sh wrapper for source-tree work:

git clone [e-mail]:amirrouh/inferhost.git
cd inferhost
./run.sh install     # creates venv, installs in editable mode
./run.sh start       # launches the TUI (downloads binaries on first run)
./run.sh status      # headless status print
./run.sh stop        # stop daemons
./run.sh test        # run pytest

Run ./run.sh help for the full list. End users do not need run.sh — they only ever type inferhost.

License

Apache 2.0