Getting Started

This quickstart shows how to run a minimal local setup from the compute-all repository: one orchestrator and one or two Physical Nodes (PNodes) on your machine, submit a network, and see DCNR in action.

Prerequisites

  • Bazel (Bazelisk recommended) — the primary build system (plain go build also works)
  • Go 1.24+ (managed by Bazel)
  • Docker (optional, for containerized runs; images are built with Bazel)
  • An S3-compatible store for VNode state (MinIO works locally)
git clone https://github.com/Free-LLM/compute-all
cd compute-all
bazel build //...        # build everything
bazel test //...         # run the test suite

1) Start the Orchestrator

The orchestrator manages network topology, PNode registration, allocation, and training sessions.

bazel run //orchestrator/cmd/orchestrator

Key environment variables: ORCHESTRATOR_PORT (default 50052) and the DynamoDB table names used for persistence.

See also: Orchestrator · Protocol

2) Start a Physical Node (PNode)

The PNode provides the compute resources and hosts the Virtual Nodes (VNodes).

ORCHESTRATOR_ADDRESS=localhost:50052 PNODE_PORT=50051 \
  bazel run //pnode/cmd/pnode

Expected logs:

  • PNode server listening at :50051
  • Successfully registered with orchestrator

Start a second PNode on another port to see real distribution.

See also: Physical Node (PNode)

3) Submit a Network

A network is a graph of typed VNodes defined in YAML. The repository ships working examples in data/ — from a tiny test network to the production transformer topologies.

Use the dcnr CLI to create the network and drive it:

bazel run //cli/cmd/dcnr -- --help

The CLI covers network creation/inspection, training sessions, dataset and tokenizer management, and live monitoring.

4) Observe Execution

  • The orchestrator allocates VNodes to available PNodes, honoring colocation groups and locality.
  • PNodes lazily instantiate VNodes and execute forward/training requests; training hops between PNodes are asynchronous.
  • Training sessions stream step metrics; the monitoring dashboard shows charts, the live network graph, and per-PNode resources.

5) Observe Fault Tolerance

Stop one of your PNodes while a network is deployed. The orchestrator detects the missing heartbeats and re-allocates its VNodes to the remaining PNodes, which restore state from storage.

6) Next Steps

Built in the open. Owned by the community.