Ollama Models

Everything Ollama lets you do with a model, as swamp methods — and a Modelfile generator that customises a model to this host's GPU and RAM, then builds it and proves it loads. WHAT IT DOES

Drives the Ollama HTTP API at OLLAMA_HOST (local or a GPU host on the LAN): list/show/ps, pull/push/copy/delete, and run/chat/embed/unload. Its point of difference is renderModelfile: it takes a GGUF path and the model's chat template from a Hugging Face download, derives num_ctx and num_gpu from the GGUF size and the host's VRAM/RAM budget, and attaches your SYSTEM instructions and parameters — so the same model is built correctly on a 4090 box, a Jetson or a Mac. create then builds the Ollama model, and verify runs it and asserts it actually loaded on the GPU (size_vram > 0) rather than silently spilling to CPU, recording tokens/s. create imports a GGUF by uploading it to Ollama as a blob (so it works when the daemon runs as another user or on a LAN host), sends numeric parameters as numbers, and omits TEMPLATE unless you supply one — Hugging Face Jinja chat templates are not Ollama Go-templates, so Ollama's own build would otherwise fail on bos_token; Ollama then uses the chat template embedded in the GGUF. Prefer it over hand-writing a Modelfile and shelling out to ollama: the fit maths, the build and the proof are repeatable swamp data.

INSTALL

swamp extension pull @svendowideit/ollama-models

DEPENDENCIES

For the methods alone: none — they talk HTTP to an Ollama server. The bundled @svendowideit/ollama-model-factory workflow additionally uses @svendowideit/gpu-facts (this host's VRAM/RAM budget) and @svendowideit/huggingface (search + download), declared in dependencies: and installed by the single pull above. Install Ollama itself with @svendowideit/ollama if the host does not have it yet.

RUN

# A Modelfile customised to this host: derive num_ctx/num_gpu from the
# GGUF size and the VRAM/RAM budget, with the chat template and your
# instructions. Writes `Modelfile` next to the build.
swamp model @svendowideit/ollama-models method run renderModelfile om \
  --input modelName=coder:latest --input from=/models/coder-Q4_K_M.gguf \
  --input template="{{ .Prompt }}" --input system="You are a coding assistant." \
  --input sizeBytes=4683071232 --input contextLength=32768 \
  --input vramMiB=24564 --input ramMiB=128000 --input outputPath=Modelfile

# Build the model from the last rendered Modelfile.
swamp model @svendowideit/ollama-models method run create om \
  --input model=coder:latest

# Prove it runs and loaded on the GPU (fails if size_vram is 0). Point
# `host` at a LAN GPU box to do this on another machine.
swamp model @svendowideit/ollama-models method run verify om \
  --input model=coder:latest

# Push the built model to your namespace on ollama.com (or a registry).
swamp model @svendowideit/ollama-models method run push om \
  --input model=coder:latest --input target=sven/coder:latest

# See what is loaded right now and how much VRAM it is using.
swamp model @svendowideit/ollama-models method run ps om

# The whole loop in one run: read the host GPU (gpu-facts), recommend a
# model (huggingface), ask you if ambiguous, download, render a
# Modelfile, build and verify it. Run it on the GPU host itself.
swamp workflow run @svendowideit/ollama-model-factory \
  --input ask="a local coding assistant" --input modelName=coder:latest

CONFIGURE

Global arguments (set at creation with --global-arg, or per call with --input):

host        Ollama server base URL                (default http://127.0.0.1:11434)
registry    Default push registry host            (default: server default)
namespace   Default push namespace                (default: push as-is)
timeoutSec  Per-request HTTP timeout              (default 120)
keepAlive   Default keep_alive for inference      (default 5m)

SETUP

setup resolves and persists the server config: it reads any existing config resource back first, probes /api/version, and records whether the host is reachable. Run it after creating the model, or when host changes:

swamp model @svendowideit/ollama-models method run setup om

UPGRADE

This extension installs nothing on the host (no service, daemon or unit), so upgrade the code alone and re-run setup:

swamp extension update @svendowideit/ollama-models

WHAT IT INSTALLS

Nothing on the host — no services, daemons or triggers. It talks HTTP to an Ollama server and writes swamp resources; renderModelfile can also write a Modelfile to a path you choose. Models it creates live in Ollama's own store, not in this extension.