Getting Started¶

Install the Development Checkout¶

You need Python 3.14.10 or newer and uv.

git clone https://github.com/salut-run/salut.git
cd salut
uv pip install -e ".[dev]"

The source checkout can run the interface and control plane with uv run python -m salut. Native inference requires the complete verified v0.8.0/b11960 Apple-silicon runtime. Build the development app with uv run poe package-macos and launch the resulting app. The package task bundles the runtime from llama-runtime.toml. To use that runtime with the source checkout, set SALUT_LLAMA_APP to the absolute path of the signed app:

export SALUT_LLAMA_APP=/absolute/path/to/Salut.app
uv run python -m salut

The source checkout verifies the same app signature, manifests, and complete runtime before launch.

uv sync --locked --extra dev
uv run python -m salut

Pair Peers¶

When another Salut node appears:

1. Open Settings → Peers.

2. Confirm the expected machine and certificate fingerprint.

3. Accept the pairing request.

4. Complete the same decision on the other node.

Only mutually paired peers can receive signed controls or participate in a distributed deployment.

Load a Local GGUF¶

With the complete verified b11960 Apple-silicon runtime, open Settings → Models → Choose GGUF File and select one complete local GGUF. The UI uses a 4096-token context and one request slot. Automatic loading does not download models or accept split GGUFs or MTP drafts.

You can also load through the local API:

curl http://localhost:7258/v1/models/load \
  -H "Content-Type: application/json" \
  -d '{"model":"my-model","path":"/absolute/model.gguf","context_size":4096}'

The API accepts context sizes from 1 through 131072, subject to native memory estimation. Salut checks memory before starting the model controller. It tries the local Metal device first, then healthy paired Metal or Windows CUDA devices at the same llama.cpp commit if more capacity is needed. A node already serving as another coordinator’s worker cannot load its own model until that worker deployment stops.

Placement prioritizes capacity. Physical comparison of predicted and actual allocations, aggregate-only capacity acceptance, and measured-bandwidth placement remain pending.

Explicit Development Deployment¶

The existing development operation also accepts one immutable execution plan through peer HTTPS. The plan identifies the exact GGUF, binary hashes, selected devices, and allocation weights.

A remote operator must be paired and explicitly allowed by full certificate fingerprint in SALUT_LLAMA_OPERATOR_FINGERPRINTS on the coordinator. The empty default denies every remote operator.

uv run python -m salut.inference.operator deploy \
  --host coordinator.local \
  --coordinator-fingerprint <full-fingerprint> \
  --plan plan.json

The same command provides status and stop operations. This control uses signed peer HTTPS. It does not require SSH.

Use the Local API¶

List the deployments visible to this API node:

curl http://localhost:7258/v1/models

Use an exact model ID returned by that operation:

curl http://localhost:7258/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<model-id-from-v1-models>",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from openai import OpenAI

client = OpenAI(base_url="http://localhost:7258/v1", api_key="unused")
response = client.chat.completions.create(
    model="<model-id-from-v1-models>",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

Next Steps¶

• Read Peers & Trust for the pairing model.

• Read Network for native multi-node execution.

• Read Models for the exact GGUF workflow.

• See the API reference for current endpoints.

• Check Release Status before treating an artifact as supported.