Skip to main content
Every endpoint page tells you where the deployment is, what its machine is doing right now, and, when something goes wrong, what failed and what to try next. The same information is available to agents through the MCP tools get_deployment_status and diagnose_deployment, which also say why a deployment is not running (stop_reason: schedule, idle, billing, spend_cap, user, failure, preempted or paused_limit) and when it next starts by itself (next_scheduled_start_at).

​
Statuses

​
What the machine is doing

While a deployment starts, its page shows live detail from the machine:
  • What the instance is doing: the current step, and while the model downloads, Downloaded (for example, 12 GB of 30 GB), the Download rate, and the Time left. If the machine hasn’t reported for more than 10 minutes, a note says the reading may be out of date.
  • Boot timeline: each start-up step the deployment went through, with its time.
  • Last provider reading: the machine’s actual state, the state Valkyrie expects, its uptime, and its hourly cost.
If a model download makes no progress for 5 minutes while the machine is still reporting, Valkyrie fails the deployment with The model download stalled, releases the machine so you stop paying for it, and avoids that machine for the next attempts. Deploy again to start on a different machine.
When a failed machine is removed, Valkyrie first captures what it can: the final logs and the machine’s last state. That record stays on the page under Captured before the failed instance was removed, so you can still read why it failed after the GPU is gone.

​
Failure codes and next steps

A failed deployment shows Deployment failed: with a short label, the error message, and a Next step. The API and MCP return the same as failure.code, failure.label, failure.summary, and failure.next_step. Some requests are refused before any GPU is rented, so they cost nothing:
  • A model name that isn’t a Hugging Face owner/name id. See Model names.
  • A disk smaller than the model’s weights. See Disk size.
  • A model whose repository files or weight format the serving engine can’t load. The error starts with Runtime preflight failed.
  • No machine that meets Valkyrie’s reliability checks for the GPU you picked. The error says how many machines each check removed. Try again later or pick another GPU type.

​
Diagnose a deployment

The Diagnosis card on every endpoint page puts the whole story in one place:
  • A headline: Serving, what the deployment is doing now, or Failed: with the failure label.
  • The next step, when there is one.
  • The latest report from the machine, the last log output, the last start-up report from the model server, the last provider reading, how long it has been in the current stage, and whether the final logs were captured before teardown.
  • The tail of the stored logs, or No logs stored yet.
Start here when a deployment misbehaves. If you contact support, include the deployment id and what this card says. From an agent, diagnose_deployment returns the same information plus up to 300 lines of stored logs. It only reads stored data and changes nothing. See MCP tools.