Skip to main content

​
Description

Deploy the current workspace, or activate an existing release. With no positional argument, this is the normal deployment path:
  1. reads the model and release from reactor.yaml,
  2. registers the model when it does not exist yet,
  3. publishes the release if it does not already have an image,
  4. applies the current model configuration, and
  5. activates the release using deployment.yaml when present.
After reactor init, the usual flow is one command:
To publish a prebuilt image instead of building one locally:
An explicit <model>:<release> skips reactor.yaml publication and configuration updates, then activates an already-published release. A local deployment.yaml still controls its regional instance plan when present. Use that form for rollbacks and CI flows that publish separately; it cannot be combined with --source. Use this command to:
  • Activate a freshly published release when it is ready for traffic.
  • Roll back to an older release if the current one has issues.
Weights: If the release has no weights of its own, the deploy reuses the most recently deployed release’s weights (server-side, no upload). When neither the release nor any previously deployed release has weights, the deploy still succeeds and prints a warning to upload them with reactor weights upload. When the workspace deploy publishes the release, the publish step reads runtime.weights_path from reactor.yaml the way reactor model publish does. --weights <path> (or an hf://org/repo[@rev] reference) names a different source for this release; --no-weights skips the step, for a machine that holds no weights. Build flags: When the workspace deploy builds the release, --build-env and --build-secret pass to that build with the same grammar as reactor model publish. An already-published release skips the build, so they have no effect. Workspace-aware: When run from a workspace (any directory under one containing a reactor.yaml, stopping at the enclosing repository root) with no positional argument, the model name, release tag, mutable model configuration, and image build are taken from the workspace. An explicit <model>:<release> activates that existing release without rebuilding it or applying the workspace spec. Its <model> part must match the workspace model.name or the command is rejected. Strategy: By default a deploy replaces instances one at a time, so the model keeps serving throughout. --strategy replace-all signals the whole old generation at once: an idle fleet turns over in seconds, and a busy one runs at reduced capacity until the fresh instances finish loading. Under both, an instance still serving a session keeps serving it until the session ends or the instance’s grace period runs out, which is up to five minutes in production. Immediate (emergency cutover): --immediate cuts that tail. After signalling the rollout, the platform disconnects the model’s live sessions so the new release takes over in seconds instead of minutes. Reach for it when the running release must stop serving now (a bad version, a safety problem), not for routine deploys. It implies --strategy replace-all (spelling the strategy out is fine; any other strategy is an error). It terminates whatever sessions are live when the sweep runs, including ones that started during the deploy. Two overlapping deploys, or a retry, each disconnect the then-live set. Clients are told the model was redeployed where their session’s release supports it; on older releases they see a plain disconnect. Their sessions end either way. --yes skips the confirmation prompt for scripts and CI. Examples:

​
Options

​
Global options

​
See also