Skip to content

Connect compute

The Compute page is where you wire up cloud GPU providers, watch what’s running, and set scaling policies. This page walks through the onboarding — the tabs, connecting a provider, and browsing the instance types a backend offers. For the conceptual overview of why and when to connect a provider, see Compute.

In-app path: Develop → Compute (/studio/compute, which lands on the Cloud tab).

The Compute header (/studio/compute) carries a tab bar with two tabs:

Tab Path What it shows
Cloud /studio/compute/cloud Connected providers, active cloud instances, cloud spend, and the provider connect grid.
Workload Queue /studio/compute/workload-queue Everything currently consuming compute — training, serving, eval, tasks, synthetic.

A third view, Policies, lives at /studio/compute/policies (auto-scaling & resource policies). It isn’t in the tab bar’s two buttons, but it’s a real page in the Compute section.

The Cloud tab stacks a few cards:

  1. Connected providers — one row per connected backend. Each row shows the provider name and tagline, a status dot (green healthy, amber degraded), the instance count (or host count for SSH fleets), the hourly cost (or GPU count for SSH), and a Remove link. Removing a cloud backend disconnects it from the project and terminates its running instances; removing an SSH fleet only drops the configuration (no hosts are touched).
  2. Active cloud instances — one row per running cloud instance (Kubernetes instances are excluded), showing the workload name, the GPU · instance type, when it started, $/hr, and a Terminate button.
  3. Cloud spend — a Daily cloud spend bar chart over the last 30 days plus By job type and By project breakdown bars. Only shown once at least one provider is connected and spend was recorded.
  4. Connect a provider — the grid of provider cards (see below). When no providers are connected yet, the tab shows just this grid with an empty-state note.

A live table of everything running on your behalf — not a mockup. It merges four real backend sources, scoped to the active project:

  • Training runs → shown as SFT (SFT/DPO) or GRPO (GRPO/PPO).
  • Models → shown as Serving.
  • Synthetic runs → shown as Synthetic.
  • Managed jobs (/api/compute/jobs) → shown as Task, including KB-compile (kb_compile) and synthetic-data pipeline jobs; the TYPE/method reads the underlying job kind (e.g. kb compile).

Rows sort active-first, then newest. Filter pills across the top narrow by type, each with a live count:

ALL · SFT · GRPO · SERVING · EVAL · TASK · SYNTHETIC

The table columns are:

Column Meaning
PRI Queue position. Only queueable workloads (training, eval, synthetic) get a number; serving and tasks read .
WORKLOAD The run / deployment name.
TYPE Training, serving, eval, task, or synthetic.
STATUS With a colored dot (running, queued, provisioning, etc.).
GPU e.g. 4× H100.
LOCATION Where it’s running (local cluster or a cloud backend).
PROGRESS / ETA Time estimate.
REQUESTED BY Who launched it.

Click any row to open the workload detail (logs, metrics). Opening a workload replaces the header with the detail view.

The Policies view (/studio/compute/policies) lists auto-scaling & resource policies. A + New Policy button sits in the header. Each policy is a card with:

  • An enable/disable Switch and the policy name.
  • A “Triggered N times · last: …” line.
  • An Edit button.
  • A When (trigger condition) / Then (action) pair.
  • Optional max spend and cooldown values.

Click a card in the Cloud tab’s Connect a provider grid. That opens the connect page at /studio/connect-cloud?provider=<key>.

Nine backends are offered:

Provider Tagline
AWS Broadest GPU instance selection
Azure Enterprise-grade GPU VMs
GCP TPUs & fast networking
RunPod Per-second billing GPU pods
Modal Serverless GPUs, scale to zero
Nebius EU-based NVIDIA GPU cloud
Oracle Cloud (OCI) Bare-metal RDMA GPU clusters
Vast.ai Cheapest GPU marketplace
SSH / On-prem Bring your own GPU servers

On the connect page you’ll see:

  • The provider’s icon, name, and tagline, with a Change link that jumps back to the connect grid.
  • A collapsible How to connect section with provider-specific setup instructions.
  • For every provider except SSH, an amber note: “Connecting lets the platform spin up GPU billed to your account at this provider.”
  • The credential form (below). CONNECT verifies the connection (button reads “Verifying connection…”) before showing a success screen.
Provider Fields
AWS Access Key ID · Secret Access Key
Azure Tenant ID · Subscription ID · Client ID (Application ID) · Client Secret
GCP GCP Project ID · Service Account Key (JSON file — the project ID auto-fills from the key if left empty)
RunPod API Key · Include Community Cloud toggle (adds cheaper, less reliable Community Cloud offers alongside Secure Cloud)
Modal Token ID · Token Secret · Environment (defaults to main)
Nebius Service Account ID · Public Key ID · Private key .pem file
Oracle Cloud (OCI) User OCID · Tenancy OCID · Fingerprint · Home Region · Private Key (.pem file)
Vast.ai API Key
SSH / On-prem SSH user (default ubuntu) · SSH port (default 22) · Private key (pasted) · Host · optional bastion / proxy jump (hostname, port, user, key)

The SSH / On-prem form lists host requirements up front (Docker pre-installed; CUDA 12.1 + NVIDIA Container Toolkit on GPU hosts; passwordless sudo; AllowTcpForwarding yes; firewall allowing SSH). On connect it probes each host and reports whether it’s reachable and what GPUs it found (e.g. 2× H100 (80GB) · 160GB total), so you can confirm before running workloads. Editing a saved fleet (?fleet_id=) pre-fills everything except the private key (leave blank to keep the saved key).

Backend offers — browsing instance types & prices

Section titled “Backend offers — browsing instance types & prices”

Once a backend is connected, you can browse the instance types it offers at /studio/backend-offers?backend=<key>. The page title reads “{Provider} — Available Instances”.

Click Fetch available instances to load the offer list, then pick a row. The offer table has columns:

Column Meaning
Instance The instance / machine type.
GPU N× <GPU name> with GPU memory, or for CPU-only.
vCPUs vCPU count.
RAM System memory.
Region Cloud region.
$/hr Hourly price, with a green spot badge on spot offers.

A Spot instances checkbox filters the list to spot offers (cheaper, but may be terminated anytime). The picked offer is echoed above the table (GPU, instance, region, price).

Modal is the exception: instead of the generic offer table it shows a Modal GPU picker where you choose a GPU spec and save it as the backend’s default.

  • Compute — when to connect a provider vs. use the bundled compute, and what you can launch on connected backends.
  • Training — launching training runs on connected compute.
  • Models — serving fine-tuned models.