Marigold

A typed inference protocol over neural network models. Typed operations -- a capability class, model, and input set -- produce immutable results which can be chained together into workflows.

Applications are defined as simple python scripts over typed-model operations. These applications are distributed with the model requirements.

Models are hosted and executed locally behind a typed inference API, orchestrated via Docker Compose.

The model set covers:

• text and image embedding,
• instruction-following (chat),
• text-to-speech,
• image generation,
• depth estimation,
• image segmentation, and
• a suite of eval models for text and image quality scoring. All models run from a shared model weight cache, independent of which application package is currently running.

For a full walkthrough with a worked example, see the setup tutorial. What follows here is the reference version.

Status

Early stage. Interfaces, config format, and the CLI may change without notice. Issues and questions are welcome; for a pull request, open an issue first to discuss the change before writing code.

Prerequisites

• Docker and Docker Compose
• NVIDIA Container Toolkit (for the GPU worker service)
• Python 3, for local development against this repo (not required just to run it -- see Getting started)
• A HuggingFace token, if you plan to use a gated model.

Getting started

pip install bayis-marigold
git clone https://github.com/bayinfosys/marigold-examples
marigold deployment start marigold-examples/chat

marigold-examples are ready-to-run application packages -- each one a directory with its own models.yaml (which models to load) and marigold.toml (how to run it: which compose services, which catalogue files, any environment variables other services in the stack need). marigold deployment start brings up the full stack -- API, worker, shared model cache, database -- configured for whichever package you point it at.

marigold deployment stop marigold-examples/chat     # tear down
marigold deployment logs marigold-examples/chat     # tail logs
marigold deployment status marigold-examples/chat   # container state
marigold cache inspect                              # what's cached, where, disk usage

Only one deployment runs at a time.

Configuration

A system-level config.toml (in the current directory, ~/.marigold/config.toml, or wherever $MARIGOLD_CONFIG points) sets host defaults. Application package marigold.toml declare what that package needs.

# config.toml -- host-level, applies to every package unless overridden
[cache]
dir = "/data/marigold"           # default: ~/.marigold/cache

[database]
url = "postgresql://..."         # default: the compose-managed Postgres

# marigold-examples/chat/marigold.toml -- per-package
[deployment]
compose_files = ["core", "webui"]
models_yaml = ["models.yaml"]

[environment]
RAG_EMBEDDING_MODEL = "sentence-transformers/all-minilm-l6-v2"

[environment] forwards arbitrary variables to non-core services in the stack -- RAG_EMBEDDING_MODEL for open-webui in this example. Marigold's own CLI doesn't interpret these; it just passes them through.

Cache location and database connection are host-level concerns.

Managing the shared model cache

Independent of any deployment:

marigold cache populate <models.yaml...>    # download missing models
marigold cache inspect                      # list what's cached, sizes, location

Multiple models.yaml files can be given together; their union is downloaded, and anything already cached by any package is skipped rather than re-fetched:

marigold cache populate marigold-examples/quick-platform-test/models.yaml \
                        marigold-examples/simple-rag/models.yaml

--prune on populate identifies models in the cache which are no longer declared anywhere.

Architecture

Marigold's Docker Compose setup is split across three files, packaged under compose/:

• docker-compose.core.yaml -- the actual services: Postgres (queue tables, plus LISTEN/NOTIFY for pub/sub), cache-init (a one-shot container that populates the model cache before anything else starts), a worker service that polls its assigned queues and runs inference, and the API service (FastAPI via uvicorn).
• docker-compose.webui.yaml -- open-webui, as an example consumer of the OpenAI-compatible endpoint. Included by a package's marigold.toml when it lists webui under compose_files.
• docker-compose.yaml -- includes both, for direct docker compose use outside the marigold CLI.

flowchart TD
    classDef svc fill:#B5D4F4,stroke:#185FA5,color:#042C53,stroke-width:1.5px
    classDef cp  fill:#FAC775,stroke:#854F0B,color:#412402,stroke-width:1.5px
    classDef cli fill:#D3D1C7,stroke:#5F5E5A,color:#2C2C2A,stroke-width:1.5px
    classDef q   fill:#9FE1CB,stroke:#0F6E56,color:#04342C,stroke-width:1.5px

    CLI([Client]):::cli
    WEBUI["open-webui\n(example consumer)"]:::cli

    PKG[/"MARIGOLD_PACKAGE_DIR\nmodels.yaml, read-only mount"/]:::svc
    INIT["cache-init\n(one-shot, populates cache)"]:::svc
    API["api service\nFastAPI / uvicorn"]:::svc
    PG[("Postgres\nqueue tables + LISTEN/NOTIFY")]:::q
    WRK["worker service\nQueueWorker / MultiQueueWorker"]:::cp
    CACHE[/"MARIGOLD_CACHE_DIR/data/models\nshared across packages"/]:::svc

    PKG -->|declares models| INIT
    INIT -->|populates| CACHE
    CLI -->|"POST /{mode}/{task}"| API
    WEBUI -->|"OpenAI-compatible API"| API
    API -->|enqueue| PG
    PG -->|"SELECT FOR UPDATE SKIP LOCKED"| WRK
    WRK -->|load weights| CACHE
    WRK -->|write result| PG
    CLI -->|"GET /{mode}/{task}/{message_id}"| API
    API -->|read status + result| PG

Inference flow

sequenceDiagram
    participant C as API client
    participant A as api service
    participant P as Postgres
    participant W as worker service
    participant M as MARIGOLD_CACHE_DIR/data/models (read-only mount)

    C->>A: POST /{mode}/{task}
    A->>P: check results cache (cache hit returns immediately)
    A->>P: write status=queued, enqueue job
    A-->>C: 200 {message_id}

    W->>P: SELECT FOR UPDATE SKIP LOCKED (poll)
    W->>M: load model weights (read-only mount, once per idle period)
    Note over W: run inference
    W->>P: write status=complete + inline result

    C->>A: GET /{mode}/{task}/{message_id}
    A->>P: read status + result
    A-->>C: 200 {status, result}

Text and vector outputs are written to and read from Postgres directly. Binary outputs (images, audio, depth maps) are written to local disk under MARIGOLD_CACHE_DIR/data/outputs.

Model catalogue files

Model weights can take a long time to download and use significant disk space. Catalogue files are kept small and task-specific rather than one large registry -- each application package's own models.yaml, declared in that package's marigold.toml under models_yaml. Multiple files can be listed together, so combining a small instruct model with an embedding model for RAG is a two-item list, not a new registry:

[deployment]
models_yaml = ["models-instruct.yaml", "models-embed.yaml"]

Model types

Type	Input	Output	Example models
text-embedding	text	vector	all-minilm-l6-v2, bge-small-en-v1.5
image-embedding	image	vector	clip-ViT-B-32
instruct	chat	chat	qwen2-1.5b-instruct, qwen2.5-3b-instruct
tts	text	audio (mp3)	mms-tts-eng, mms-tts-cym
txt2img	text	image (png)	stable-diffusion-3.5-large-turbo, flux.1-schnell
img2txt	image	text	paligemma2
depth	image	depth map (png)	dpt-dinov2-small-kitti
img2mask	image	segmentation mask (png)	sam-vit-huge
text-eval	text	scores	toxic-bert, distilbert-sst2
text-similarity	text pair	similarity score	all-minilm-l6-v2
image-eval	image	scores	nsfw-image-detection, cafe-aesthetic
image-text-eval	image + text	alignment score	clip-ViT-B-32

image-embedding has an implemented handler but no tested catalogue entry as of writing.

OpenAI-compatible endpoints

GET /v1/models, POST /v1/chat/completions, POST /v1/embeddings are implemented as a synchronous wrapper around Marigold's async submit/poll path, for use with unmodified OpenAI-SDK clients (open-webui, LangChain, LangGraph, Continue.dev, and others).

Two limitations worth knowing before relying on this for anything beyond basic chat:

• stream=true is accepted and returns correctly-framed SSE chunks, but there is no token-level streaming.
• Only one tool call per assistant turn round-trips correctly. The wire format has no field for tool_call_id, so a turn with parallel tool calls cannot be matched back to the call that produced it.

Repository structure

package/src/
  api/                            -- API definitions; routes/ is a package
  models/                         -- model handler code (one file per model type)
  shared/                         -- enums, registry, output persistence, usage tracking
  cli/
    main.py                       -- the `marigold` command
  compose/
    Dockerfile                    -- consolidated multi-stage build (api / cache / worker-cpu / worker-gpu)
    docker-compose.core.yaml
    docker-compose.webui.yaml
    environment/                  -- per-target requirements files
  tools/
    model_cache_shared.py         -- cache build/inspect logic
    model_cli.py                  -- model, cache, workflow, and status CLI (runs inside cache-init/worker)

marigold-examples (application packages -- models.yaml + marigold.toml + optional scripts) live in a separate repository.

Handler architecture

Registry and decorator

Every model type is registered at import time using the @model_spec decorator from shared.registry. The decorator populates the _SPECS singleton dict, keyed by ModelType.value. A ModelSpec instance couples: the ModelType enum value, the ModelMode (embed / eval / gen), the loader function, the handler class implementing _run(), the request and response Pydantic models, the binary OutputField declarations, and the API route path.

Loader contract

Every loader function must return a ModelLoaderResult:

@dataclass
class ModelLoaderResult:
    processor: Any   # tokenizer, image processor, or None
    model: Any       # the model, pipeline, or SentenceTransformer

standard_loader in models/standard_loader.py handles the common AutoTokenizer / AutoProcessor + AutoModel pattern. Model-type-specific loaders select the correct transformer classes and delegate to it.

Handler contract

BaseModelHandler.process() validates the request dict against ModelSpec.request_model and calls self._run() with the typed result. Subclasses implement only _run().

Adding a model

1. Add an entry to the relevant package's models.yaml.

2. If the type is new, create a handler file in package/src/models/ following the existing pattern; if the type exists, the model is served by the existing handler.

3. Register the new handler import in models/load_all().

   Checks the schema and flags duplicate (name, type) entries. Then restart against the new catalogue to verify the model caches and loads correctly.

Authentication

No API key is required. Requests are accepted from any caller; the caller is identified by an optional X-User-Id header, defaulting to local-user (see auth.py, get_authorizer()).

This is designed for use on localhost or within a private network. Marigold does not provide its own authentication layer -- if you expose it beyond that boundary, put your own auth in front of it.