> ## Documentation Index
> Fetch the complete documentation index at: https://docs.scitix.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Prepare Instance Resources

> Before creating a development environment, training task, or inference service, confirm the available GPU/CPU instance resources, resource pools, and quotas.

Instance resources provide the GPU/CPU computing power used by development environments, training tasks, and inference services. Before creating a workload, you need to understand the resource pool types, purchase or verify the instance quota, and determine which resource pool the quota will be assigned to. Quotas purchased without a specified target resource pool will be assigned to the shared resource pool and can be adjusted later as needed.

## Applicable Scenarios

* Create JupyterLab or VS Code Server development environments.
* Submit training jobs such as PyTorchJob and RayJob.
* Publish LLM inference services or general services.
* Reserve a fixed amount of computing capacity for your team.

## Learn About Resource Pool Types

When selecting a resource pool, first determine whether your task requires dedicated guarantees or node isolation.

| Resource Pool Type | Recommended Scenarios | Points to Note |
| - | - | - |
| shared resource pool | General development, short-term debugging, elastic tasks, or tasks that do not require guaranteed fixed resources. | Resources are shared among users within the tenant, and tasks may be affected by overall usage. You can manage authorized users and user-level configurations on the details page. |
| dedicated resource pool | Tasks that require fixed resources to be allocated to specified users or teams, or that restrict certain workload types to specific resources. | Pay attention to the total resource pool quota, authorized users, individual quotas, and permitted workload types. |
| exclusive resource pool | Tasks requiring exclusive access to an entire node, resource isolation, or node health monitoring. | The quota must meet the full-node validation requirements. Defragmentation may be required after long-term use. |

## Buy Instance

If you do not have sufficient instance quota, you can first purchase instance resources. Go to **Resources > Order**, and under the **Instance** tab, click **Buy**.

When purchasing an instance, select the Region and Cluster to which the resource belongs, and configure the following parameters:

| Parameter | Description |
| - | - |
| Instance | The purchased instance specification, which must meet the GPU, CPU, and memory requirements of the task. |
| Quantity | The amount of quota to purchase. |
| Runtime | The duration for which the resources are purchased, in days. |
| Target Resource Pool | Specify which resource pool the quota will be added to. If not specified, it will be added to the shared resource pool. |
| Auto-renewal | Whether to automatically renew upon expiration. |

When purchasing instance resources, pay close attention to **Target Pool**. It determines which resource pool the instance quota purchased in this transaction will be allocated to:

| Quota Destination | Applicable Scenario | How to Proceed |
| - | - | - |
| shared resource pool | Resources are shared by users within the tenant, or no dedicated resource pool has been planned yet. | Do not specify a target resource pool when purchasing. The instance quota is allocated to the shared resource pool by default. |
| Existing dedicated resource pool | A fixed quota needs to be assigned to specified users or teams. | Specify the target dedicated resource pool when purchasing. |
| Existing exclusive resource pool | Whole-node exclusivity, resource isolation, or node health monitoring is required. | Specify the target exclusive resource pool when purchasing, and ensure that the purchase quantity meets the whole-machine validation requirements: GPU instances must be aligned to 8 GPUs per machine, and CPU instances must be aligned to 120 CUs per machine. |
| New dedicated or exclusive resource pool | No target resource pool currently meets the requirements. | You can first create a resource pool, then purchase and specify the target resource pool. Alternatively, you can first purchase resources into the shared resource pool and then allocate them through resource pool scaling. For detailed instructions on creating dedicated or exclusive resource pools, see [Manage Resource Pools and Quotas](/siverse/resources/manage-resource-pools-and-quotas.md). |

## Estimate Task Resource Requirements

Before submitting a development environment, training task, or inference service, estimate the required resources based on the instance specification, number of specifications, and number of replicas, then verify the target resource pool and your remaining quota.

**Total task requirements = Resources per instance specification × Instance specifications per replica × Number of replicas.**

For example, an instance specification includes 2 GPUs, 16 vCPUs, and 128 GiB of memory. Each replica in the task uses 3 such specifications, and 4 replicas are running in total. The resource requirements are as follows:

| Resource | Per Specification | Single Replica (× 3 specifications) | Entire Job (× 4 replicas) |
| - | - | - | - |
| GPU | 2 | 6 | 24 |
| CPU | 16 vCPU | 48 vCPU | 192 vCPU |
| Memory | 128 Gi | 384 Gi | 1536 Gi |

This task requires a total of 12 units of the specified configuration. Before submitting, verify this total against **Remaining Quota** and the available capacity in the target resource pool for the selected instance specification.

The computation methods for different task types must also take role differences into account:

* The total number of replicas for a PyTorch job includes the master node and Workers; the master node also consumes resources.
* For a Ray job, calculate the resource requirements for the Head and each type of Worker separately. Aggregate and verify identical specifications; verify quotas separately for different specifications.
* Different GPU models must be checked separately for model and capacity. The number of idle cards of one GPU model cannot offset the resource requirements of another model.

## Check Whether Resources Are Available

After completing the purchase, first verify that the order status is `active` in Orders, then return to Instance or the target resource pool to confirm that the quota has increased.

Go to **Resources > Instance**, then select a region, cluster, and project group. The resource overview summarizes resources within the current scope, including:

| Metrics | Description |
| - | - |
| Instance | Used quantity, total quantity, and utilization of the current instance resources. |
| CPU Cores | Used quantity, total quantity, and utilization of the current CPU cores. |
| GPU Cards | Used quantity, total quantity, and utilization of the current GPU cards. |
| Memory | Used capacity, total capacity, and utilization of the current memory. |
| Number of Resource Pools | Number of shared, dedicated, and exclusive resource pools in the current region and cluster. |

In the resource pool list, continue checking whether the target resource pool meets the task requirements. The list supports custom display columns. Commonly used columns include:

| Information | Purpose |
| - | - |
| Pool Name / Description | Identify the target resource pool and confirm that the expected resource pool has been selected. |
| Spec / Billing | Determine whether the resources match the task configuration and cost expectations. |
| Instance Quota | View the total quota available for allocation in the resource pool. |
| Used / Idle / Unavailable | Determine whether resources are still available for task submission. |
| Tasks (Running / Queued) | Determine whether resources are occupied by existing tasks or whether tasks are waiting to be scheduled. |
| Status | Determine whether the resource pool is available. |
| Personal Quota | Determine the amount of personal resources available to you in the current resource pool. |
| Workload Types | Determine whether development environments, training tasks, or inference services are allowed to use the resource pool. |
| Frozen | Determine whether the resource pool is frozen. |
| Idle-Time Sharing | Check whether idle-time sharing is enabled for the resource pool. |
| Resource Guarantee | Check whether resource guarantee is configured for the resource pool. |
| Default User Priority | View the default scheduling priority used when users submit tasks. |
| User Default Quota | Determine the personal quota granted by default to newly authorized users. |
| Project Group | Determine the project group to which the resource pool belongs. |
| Administrator | Confirm the users who can manage the resource pool. |
| Member Count | View the number of users authorized for the current resource pool. |

If the utilization of a certain type of resource is high, or the target resource pool does not have sufficient available quota, subsequent task creation may be queued or submission may fail. If the resource pool is being updated, deleted, or is in a failed state, wait for the operation to complete or contact the resource administrator for confirmation before submitting the task.

## Request Resource Access Permissions and Quotas

If the target resource pool has not been authorized for the current account, or your remaining personal quota is insufficient, first apply for permission to use the resource pool or request additional quota.

A resource request goes through two stages: approval and activation. After an administrator approves the request, the system must still apply the authorization and quota configuration. Only after the request status changes to **Active** can the newly granted permissions or quotas be used to create Develop environments, Training jobs, or Inference services.

1. Go to **Resources > Instance**.
2. Select the target region, cluster, and project group, then locate the target resource pool.
3. Verify the resource pool’s instance specification, supported workload types, status, and personal quota.
4. If **Request Access** is available, enter the required quota and purpose, then submit the request. If no request option is available, contact the resource pool administrator.
5. View the approval and activation progress in the request records.
6. After the status changes to **Effective**, refresh the resource pool page or the task creation form to confirm that your remaining personal quota has been updated.

Application status descriptions are as follows:

| Status | Description | Recommended Action |
| - | - | - |
| Pending (`pending`) | The request has been submitted and is awaiting administrator review. | Wait for approval. If changes are needed, withdraw the request as permitted on the page and resubmit it. |
| Approved (`approved`) | The request has been approved, and the system is provisioning the authorization and quota. | Wait for it to take effect. |
| Effective (`succeeded`) | The authorization and quota have been configured. | Refresh the page and verify the actual quota. |
| Rejected (`rejected`) | The request did not pass the review. | Check the rejection reason, revise the requirements, or contact an administrator. |
| Withdrawn (`cancelled`) | The request has been cancelled. | If you still need the resources, submit a new request. |
| Activation Failed (`failed`) | The request was approved, but configuration was not completed. | Provide the request ID and error reason to the administrator. |

Approval does not mean that resources are immediately available. After authorization and quota take effect, tasks must still wait for schedulable resources that meet the requirements. If a task remains queued for an extended period, see [Troubleshoot insufficient resources or long task queue times](/siverse/resources/troubleshoot-insufficient-resources-or-queued-tasks.md).

## Related Documentation

* [Manage resource pools and quotas](/siverse/resources/manage-resource-pools-and-quotas.md)
* [View and renew resource orders](/siverse/resources/manage-resource-orders.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.