Instances

These are the endpoints the console calls when you launch, watch and destroy an instance. They accept only a browser session, not an API key. To script instances, use Radeon Cloud CLI.

Paths are relative to your site’s console address, for example https://developer.amd.com.cn/radeon. Each account can have one instance at a time, so none of these endpoints take an instance ID.

Task	Notebook	Model service
Launch	POST /api/notebook/request	POST /api/inference or POST /api/templates/{id}/launch
Check status	GET /api/notebook/status	GET /api/inference/current
Read logs	GET /api/notebook/logs	GET /api/inference/current/logs
Destroy	DELETE /api/notebook/current	DELETE /api/inference/current

Choose resources

A launch names a resource pool and a size. Both come from GET /api/me/resource-labels, which lists what your account may use:

• entries with "category": "resource-pool" are pools; send the id as pod_type
• entries with "category": "resource-template" are sizes; send the id as resource_template

For example, pod_type: "one-click" with resource_template: "resource-16c55g1u" is 1 GPU, 16 cores and 55 GB of memory. Accounts that have pools and sizes assigned must send both fields and must not send gpu_count.

Launch a notebook

POST /api/notebook/request Session

Field	Type		Description
pod_type	string	Required	Resource pool ID.
resource_template	string	Required	Size ID.
image	string	Optional	An image from the catalog. Defaults to the platform default, AMD OneClick Base.
instance_type	string	Optional	Must be jupyter, the default.
use_pvc	boolean	Optional	true mounts your persistent /workspace. Default false, a temporary workspace.
pvc_size	string	Optional	25GB or 100GB, with use_pvc: true. 100 GB is offered only to some accounts.
model_mount	string	Optional	A model to mount read-only. GET /api/models/mountable lists the IDs, each with the mount_path it appears at, usually /models.
ssh_enabled	boolean	Optional	The console always sends true. Requires an SSH key in Settings.

{
  "pod_type": "one-click",
  "resource_template": "resource-16c55g1u",
  "instance_type": "jupyter",
  "use_pvc": true,
  "pvc_size": "25GB",
  "ssh_enabled": true
}

The call returns as soon as the launch is accepted; it doesn’t wait for the instance. Poll status until ready is true.

The launch is refused when:

• you already have an instance (400, Each user can only have one active instance), or the previous one is still being deleted (409)
• your credits don’t cover one hour on the chosen size (400, Insufficient credits). Only the Global site checks this up front. On the China site, the instance is deleted once an hour can’t be charged.
• instance_type isn’t jupyter (400, unsupported_deploy_type)
• the pool or size isn’t assigned to your account (403, resource_selection_not_assigned)

Launch a model service

To serve your own model (Deploy type: Model API), send the model name and the vLLM command:

POST /api/inference Session

Field	Type		Description
api_model	string	Required	The name clients send as model. Use the same value as --served-model-name in the command.
start_command	string	Required	A vllm serve command listening on port 8000.
model_source	string	Optional	huggingface (default) or modelscope.
pod_type, resource_template	string	Required	As for a notebook.

{
  "api_model": "MiniCPM5-1B",
  "model_source": "modelscope",
  "start_command": "vllm serve OpenBMB/MiniCPM5-1B --host 0.0.0.0 --port 8000 --served-model-name MiniCPM5-1B",
  "pod_type": "one-click",
  "resource_template": "resource-16c55g1u"
}

Unknown fields are rejected. To launch one of the ready-made model templates instead, list them with GET /api/inference/templates and call POST /api/templates/{id}/launch with an empty body. See Templates.

Check status

GET /api/notebook/status Session

For a model service, call GET /api/inference/current. It returns the same fields plus runtime and credit details.

{
  "status": "ready",
  "ready": true,
  "message": "Your workspace is ready",
  "instance_id": "<instance-id>",
  "instance_type": "jupyter",
  "url": "https://developer.amd.com.cn/radeon/instances/<instance-id>/",
  "ssh_host": "<ssh-host>",
  "ssh_port": 32014,
  "ssh_username": "root",
  "ssh_command": "ssh root@<ssh-host> -p 32014",
  "api_base_url": null,
  "api_key": null,
  "api_model": null
}

Field	Description
ready	true only when the instance is usable. Wait on this, not on status.
status, message	Progress or failure, for display. failed means the launch won’t recover.
phase, reason, detail, error_code	Lower-level scheduling state. Useful when a launch is stuck.
instance_id	Appears in the instance’s URLs. It can stay the same across launches, but read it from here rather than storing it.
url	Opens JupyterLab in a signed-in browser.
ssh_host, ssh_port, ssh_username, ssh_command	Filled in once the container is running.
persistent_available, storage_notice	Whether the persistent /workspace was mounted, and why not if it wasn’t.
api_base_url, api_key, api_model	Filled in for a ready model service. See Model service endpoints.

Poll every few seconds. A launch is usually ready within a minute, but a large image or a model download can take several minutes.

Read logs

GET /api/notebook/logs Session

Returns scheduling events and container output for the current instance. This is what View logs shows in the console. For a model service, call GET /api/inference/current/logs.

Destroy the instance

DELETE /api/notebook/current Session

For a model service, call DELETE /api/inference/current.

{
  "success": true,
  "message": "Instance <instance-id> destroyed",
  "destroyed_count": 1
}

If there’s nothing to destroy, the response has success: false and "No active instance found". This isn’t an error, so it’s safe to call twice.

Files in a temporary workspace are deleted with the instance. A persistent /workspace is kept. See Keep your data.