Skip to content

Remote Workers Tab

Overview

The Remote Workers tab (RemoteWorkersPage.qml, route RemoteWorkers) manages Remote SSH Workers — dedicated remote Linux/GPU hosts orchestrated by the backend to execute universal ML worker containers for training and inference jobs.

Remote SSH Workers page displaying GPU & VRAM specs, Active / Max slots, Host Profile, Readiness, Status badges, and node lifecycle actions.

Availability & Prerequisites

  • Server / Orchestrator Connection: Requires an active authenticated connection to a backend orchestrator supporting SSH compute providers (Modes 2–4).
  • Manage Permissions: Edit, Configure, Reboot, and Delete row actions are enabled only when rowData.can_manage is true for your signed-in role/tenant.

Step-by-Step User Workflows

1. Register a New SSH Worker Node

  1. Click Add Worker Node (+ primary button) in the header to open RemoteWorkerDialog (workerDialog.openForAdd()).
  2. Enter node display name, SSH hostname/IP, port, SSH user/authentication parameters, host profile, and concurrency capacity (max slots), then save.
Add Remote Worker modal with Node Display Name, Host/IP, SSH credentials, Host Key Verification, and GPU slot configuration.

2. Probe Configuration, Configure Runtime, or Edit

  • Check configuration (View icon): Calls testConnection(rowData.id) to verify SSH reachability, GPU telemetry, and readiness without mutating the node.
  • Configure: Runs idempotent remote provisioning (configureExistingNode(rowData.id)) to bootstrap Docker/worker runtime prerequisites (view streaming setup logs in Tasks).
  • Edit: Opens RemoteWorkerDialog prefilled (openForEdit(rowData)) to update node parameters.

3. Safe Reboot or Deletion

  • Reboot node (rebootConfirmDialog): Permitted only when can_manage is true, no backend busy flag is active, active_jobs === 0, and status !== "configuring".
  • Delete node (deleteConfirmDialog): Prompts "Are you sure you want to delete node <name>? Active or queued jobs targeting this node may fail." before unregistering the node.
flowchart LR
    Add["Add Worker Node (RemoteWorkerDialog)"] --> Check["Check Configuration (SSH + GPU Probe)"]
    Check --> Provision["Configure Host Runtime / Docker Worker"]
    Provision --> Ready["Status Ready · Slots: 0 / Max"]
    Ready --> Select["Selectable in TrainingJobDialog"]
    Ready -->|active_jobs == 0| Reboot["Safe Node Reboot"]

UI Controls & Table Columns Reference

Header & Polling Behavior

  • Header Title: Remote SSH Workers — "Manage remote GPU compute nodes for distributed model training and inference".
  • Automatic Polling: Starts automatically when the tab mounts (Component.onCompleted -> startPolling()) and stops cleanly when navigating away (Component.onDestruction -> stopPolling()).
  • Search & Refresh: Search bar "Search nodes by name or host..." plus manual Refresh (refreshNodes()).

Workers Table (workersTable)

Column Key Header Label Width Description
name Node Name Flex fill Friendly identifier shown in TrainingJobDialog node selector
host Host / IP 150px SSH hostname or IP address
port Port 60px SSH TCP port (typically 22)
gpu_summary GPU & VRAM 190px Detected GPU device(s) and VRAM capacity
slots Active / Max 100px Running jobs vs maximum concurrent worker slots
host_profile_display Host Profile 240px Configured runtime/environment profile for the remote host
profile_status Readiness 115px Provisioning and environment readiness state
status Status 100px (badge) Live node state badge (statusVariant), including configuring
actions Actions Row action icons Check configuration, Configure, Reboot, Edit, Delete

Technical Reference: Security & Credential Boundary

  • Backend-Owned SSH Secrets: The desktop submits node registration/updates to the FastAPI control plane; SSH keys and long-lived infrastructure credentials remain managed by the server provider layer.
  • FIFO Queuing Behavior: Inside TrainingJobDialog, nodes with active_jobs >= 1 display [<status> (Full - Queued)], indicating new jobs will wait in the node’s FIFO queue until an active slot clears.
  • Worker Isolation: Remote workers receive only versioned job envelopes, scoped control tokens, and temporary artifact URLs — never database passwords or identity-provider secrets.