truefoundry-volumes

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Routing note: For ambiguous user intents, use the shared clarification templates in references/intent-clarification.md.

路由说明：如果用户意图不明确，请使用 references/intent-clarification.md 中的通用澄清模板。

Volumes

卷

Create and manage persistent volumes on TrueFoundry. Volumes provide shared, low-latency disk storage that persists across container restarts and can be mounted by multiple pods.

在TrueFoundry上创建和管理持久卷。卷提供共享的低延迟磁盘存储，数据在容器重启后依然留存，并且可被多个Pod挂载。

When to Use

适用场景

Create, list, or mount persistent volumes on TrueFoundry, including dynamic provisioning, static PV attachment, storage class selection, and Volume Browser setup.

需要在TrueFoundry上创建、查询或挂载持久卷的场景，包括动态配置、静态PV关联、存储类选择和卷浏览器设置。

When NOT to Use

不适用场景

User needs large archival storage or global access -> suggest blob storage (S3/GCS) instead
User wants ephemeral scratch space -> use
```
ephemeral_storage
```
in resource config
User wants to deploy an app -> prefer
```
deploy
```
skill; ask if the user wants another valid path
User wants to manage secrets -> prefer
```
secrets
```
skill; ask if the user wants another valid path

</objective> <context>

用户需要大容量归档存储或全局访问权限 -> 建议使用对象存储（S3/GCS）替代
用户需要临时暂存空间 -> 在资源配置中使用
```
ephemeral_storage
```
用户想要部署应用 -> 优先使用
```
deploy
```
skill；询问用户是否需要跳转其他有效路径
用户想要管理密钥 -> 优先使用
```
secrets
```
skill；询问用户是否需要跳转其他有效路径

</objective> <context>

Volumes vs Blob Storage

卷 vs 对象存储

Help the user choose the right storage type:

Aspect	Volumes	Blob Storage (S3/GCS)
Access method	Standard file system APIs (open, read, write)	SDK clients (boto3, gcsfs)
Speed	Faster (local-disk latency)	Slower (network round-trips)
Durability	High	Extremely high (11 nines)
Cost	Higher per GB	Lower per GB
Scope	Region/cluster limited	Global access
Best for	Shared model weights, training checkpoints, low-latency reads	Large archives, datasets accessed infrequently, cross-region data

Choose volumes when:

Multiple pods need concurrent file system access to the same data
You need file system semantics (locking, renaming, directory listing)
Low-latency access to frequently-read data (model weights, config files)
ML training checkpointing where write speed matters

Choose blob storage when:

Data is larger than a few hundred GB
Global or cross-region access is needed
Data is written once and read occasionally
Cost is the primary concern

Warning: Do not write to the same file path from multiple pods simultaneously -- this can cause data corruption. Coordinate writes across pods or use separate paths.

帮助用户选择合适的存储类型：

对比项	卷	对象存储（S3/GCS）
访问方式	标准文件系统API（open、read、write）	SDK客户端（boto3、gcsfs）
访问速度	更快（本地磁盘延迟）	更慢（网络往返开销）
耐久性	高	极高（11个9）
成本	每GB成本更高	每GB成本更低
作用范围	区域/集群限制	全局访问
最佳适用场景	共享模型权重、训练检查点、低延迟读取	大型归档、低频访问数据集、跨区域数据

选择卷的场景：

多个Pod需要并发对同一份数据进行文件系统访问
需要文件系统语义（锁、重命名、目录列表）
对高频访问数据（模型权重、配置文件）有低延迟要求
对写入速度有要求的ML训练检查点存储

选择对象存储的场景：

数据大小超过几百GB
需要全局或跨区域访问
数据为一次写入、偶尔读取的类型
成本是首要考量因素

警告： 不要同时从多个Pod写入同一路径的文件 -- 这可能导致数据损坏。请协调跨Pod的写入操作，或使用独立路径。

Prerequisites

前置条件

Always verify before deploying:

Credentials --
```
TFY_BASE_URL
```
and
```
TFY_API_KEY
```
must be set (env or
```
.env
```
)
Workspace --
```
TFY_WORKSPACE_FQN
```
required. Never auto-pick. Ask the user if missing. Volumes are workspace-scoped: a volume created in one workspace can only be used by applications in that same workspace.
Cluster storage class -- The target cluster must have a storage provisioner configured for the desired storage class.

For credential check commands and .env setup, see

references/prerequisites.md

部署前务必验证：

凭证 -- 必须设置
```
TFY_BASE_URL
```
和
```
TFY_API_KEY
```
（环境变量或
```
.env
```
文件中）
工作空间 -- 需要
```
TFY_WORKSPACE_FQN
```
。绝对不要自动选择，缺失时询问用户。 卷是工作空间级别的资源：在一个工作空间创建的卷仅能被同一工作空间的应用使用。
集群存储类 -- 目标集群必须为所需的存储类配置了存储配置器。

凭证检查命令和.env设置方式请查看

references/prerequisites.md

。

Volume Types

卷类型

Dynamic Volumes (Create New)

动态卷（新建）

TrueFoundry provisions a new Kubernetes PersistentVolumeClaim (PVC). You specify size and storage class; the cluster allocator handles the rest.

Key properties:

Size is expandable after creation but cannot be reduced
Access mode is
```
ReadWriteMany
```
(multiple pods can mount simultaneously)
Reclaim policy is
```
Retain
```
(data persists even if the volume resource is deleted from TrueFoundry)

TrueFoundry会配置新的Kubernetes PersistentVolumeClaim（PVC）。你只需要指定大小和存储类，集群分配器会处理剩余工作。

核心属性：

创建后可扩容，但无法缩容
访问模式为
```
ReadWriteMany
```
（多个Pod可同时挂载）
回收策略为
```
Retain
```
（即使从TrueFoundry删除卷资源，数据依然留存）

Static Volumes (Use Existing)

静态卷（使用现有卷）

Mount a pre-existing Kubernetes PersistentVolume by name. Use this for:

AWS EFS -- Elastic File System via
```
efs.csi.aws.com
```
driver
AWS S3 -- S3 buckets via
```
s3.csi.aws.com
```
driver
GCP Filestore / GCS -- via
```
gcsfuse.csi.storage.gke.io
```
driver
Azure Files / Azure Blob -- via
```
file.csi.azure.com
```
or
```
blob.csi.azure.com
```
drivers

Static volumes require the PersistentVolume to already exist in the Kubernetes cluster. See the "Static Volume Setup" section below.

按名称挂载预先存在的Kubernetes PersistentVolume。适用场景：

AWS EFS -- 通过
```
efs.csi.aws.com
```
驱动的弹性文件系统
AWS S3 -- 通过
```
s3.csi.aws.com
```
驱动的S3存储桶
GCP Filestore / GCS -- 通过
```
gcsfuse.csi.storage.gke.io
```
驱动
Azure Files / Azure Blob -- 通过
```
file.csi.azure.com
```
或
```
blob.csi.azure.com
```
驱动

静态卷要求Kubernetes集群中已经存在对应的PersistentVolume。请查看下方「静态卷设置」章节。

Storage Classes by Cloud Provider

各云厂商存储类

For storage class tables by cloud provider (AWS, GCP, Azure) and discovery commands, see

references/volume-storage-classes.md

</context> <instructions>

各云厂商（AWS、GCP、Azure）的存储类对照表和查询命令请查看

references/volume-storage-classes.md

。

</context> <instructions>

Creating a Volume

创建卷

When using direct API, set

TFY_API_SH

to the full path of this skill's

scripts/tfy-api.sh

. See

references/tfy-api-setup.md

for paths per agent.

使用直接API时，请将

TFY_API_SH

设置为当前skill的

scripts/tfy-api.sh

完整路径。各Agent的路径请查看

references/tfy-api-setup.md

。

Before Creating

创建前须知

ALWAYS ask the user these questions in order:

Volume type -- "Do you want to create a new volume or use an existing Kubernetes PersistentVolume?"
- Create new → proceed with dynamic volume questions below
- Use existing → ask for the PersistentVolume name in Kubernetes, then skip to workspace
Volume name -- What should the volume be called?
Size -- How much storage? (integer in GB, e.g.
```
50
```
). Cannot be reduced later.
Storage class -- Which storage class? Present available options from the cluster.
Workspace -- Which workspace? Volumes are workspace-scoped.
Volume Browser -- "Do you want to enable Volume Browser? It provides a web UI to browse and manage files in your volume without SSH." (Optional)
- If yes, ask for:
  - Endpoint host -- The hostname where the browser will be accessible (e.g.
```
my-cluster.example.truefoundry.com
```
    ). Present available hosts from the cluster's base domain.
  - Endpoint path -- URL path prefix (optional, defaults to
```
/
```
    )
  - Username -- Login username for the browser (optional, defaults to
```
admin
```
    )
  - Password secret -- FQN of a TrueFoundry secret containing the browser password. If user doesn't have one, help them create it using the
```
secrets
```
    skill first.

Present a summary and ask for confirmation:

Volume to create:
  Type:          Create new (dynamic)
  Name:          training-data
  Size:          100 GB
  Storage class: efs-sc
  Workspace:     my-cluster:my-workspace
  Volume Browser: Enabled
    Endpoint:    https://my-cluster.example.truefoundry.com/training-data/
    Username:    admin
    Password:    (secret: my-cluster:my-workspace:vol-browser-pw)

Note: Size can be expanded later but not reduced.
Proceed?

For volumes without Volume Browser:

Volume to create:
  Type:          Create new (dynamic)
  Name:          training-data
  Size:          100 GB
  Storage class: efs-sc
  Workspace:     my-cluster:my-workspace
  Volume Browser: Disabled

Note: Size can be expanded later but not reduced.
Proceed?

务必按顺序向用户确认以下问题：

卷类型 -- 「你想要创建新卷，还是使用现有Kubernetes PersistentVolume？」
- 创建新卷 → 继续下方动态卷相关问题
- 使用现有卷 → 询问Kubernetes中PersistentVolume的名称，然后跳过到工作空间确认环节
卷名称 -- 卷的命名是什么？
容量 -- 需要多少存储？（GB为单位的整数，例如
```
50
```
）。后续无法缩容。
存储类 -- 使用哪个存储类？提供集群中可用的选项。
工作空间 -- 所属工作空间是哪个？卷是工作空间级别的资源。
卷浏览器 -- 「你是否需要启用卷浏览器？它提供Web UI，无需SSH即可浏览和管理卷中的文件。」（可选）
- 若选择是，询问以下信息：
  - 访问域名 -- 浏览器可访问的主机名（例如
```
my-cluster.example.truefoundry.com
```
    ）。提供集群基础域名下的可用主机。
  - 访问路径 -- URL路径前缀（可选，默认为
```
/
```
    ）
  - 用户名 -- 浏览器登录用户名（可选，默认为
```
admin
```
    ）
  - 密码密钥 -- 存储浏览器密码的TrueFoundry密钥FQN。如果用户没有对应密钥，先使用
```
secrets
```
    skill帮助用户创建。

提供信息摘要并请求确认：

待创建卷信息：
  类型:          新建（动态卷）
  名称:          training-data
  容量:          100 GB
  存储类:        efs-sc
  工作空间:      my-cluster:my-workspace
  卷浏览器:      已启用
    访问地址:    https://my-cluster.example.truefoundry.com/training-data/
    用户名:      admin
    密码:        (密钥: my-cluster:my-workspace:vol-browser-pw)

注意：容量后续可扩容但无法缩容。
是否继续？

未启用卷浏览器的卷确认模板：

待创建卷信息：
  类型:          新建（动态卷）
  名称:          training-data
  容量:          100 GB
  存储类:        efs-sc
  工作空间:      my-cluster:my-workspace
  卷浏览器:      未启用

注意：容量后续可扩容但无法缩容。
是否继续？

Via Tool Call

通过工具调用创建

tfy_applications_create_deployment(
    manifest={"type": "volume", "name": "my-volume", "config": {"type": "dynamic", "size": 100, "storage_class": "efs-sc"}},
    options={"workspace_id": "ws-id-here"}
)

For Volume Browser fields and static volume tool-call examples, use the same fields as the Direct API examples below.

tfy_applications_create_deployment(
    manifest={"type": "volume", "name": "my-volume", "config": {"type": "dynamic", "size": 100, "storage_class": "efs-sc"}},
    options={"workspace_id": "ws-id-here"}
)

卷浏览器字段和静态卷的工具调用示例，使用与下方直接API示例相同的字段即可。

Via Direct API

通过直接API创建

Create new volume (without Volume Browser):

bash

$TFY_API_SH PUT /api/svc/v1/apps '{
  "manifest": {
    "type": "volume",
    "name": "my-volume",
    "config": {
      "type": "dynamic",
      "size": 100,
      "storage_class": "efs-sc"
    }
  },
  "workspaceId": "ws-id-here"
}'

Create new volume (with Volume Browser):

bash

$TFY_API_SH PUT /api/svc/v1/apps '{
  "manifest": {
    "type": "volume",
    "name": "my-volume",
    "config": {
      "type": "dynamic",
      "size": 100,
      "storage_class": "efs-sc"
    },
    "volume_browser": {
      "username": "admin",
      "password_secret_fqn": "my-cluster:my-workspace:vol-browser-pw",
      "endpoint": {
        "host": "my-cluster.example.truefoundry.com",
        "path": "/my-volume/"
      }
    }
  },
  "workspaceId": "ws-id-here"
}'

新建卷（未启用卷浏览器）：

bash

$TFY_API_SH PUT /api/svc/v1/apps '{
  "manifest": {
    "type": "volume",
    "name": "my-volume",
    "config": {
      "type": "dynamic",
      "size": 100,
      "storage_class": "efs-sc"
    }
  },
  "workspaceId": "ws-id-here"
}'

新建卷（启用卷浏览器）：

bash

$TFY_API_SH PUT /api/svc/v1/apps '{
  "manifest": {
    "type": "volume",
    "name": "my-volume",
    "config": {
      "type": "dynamic",
      "size": 100,
      "storage_class": "efs-sc"
    },
    "volume_browser": {
      "username": "admin",
      "password_secret_fqn": "my-cluster:my-workspace:vol-browser-pw",
      "endpoint": {
        "host": "my-cluster.example.truefoundry.com",
        "path": "/my-volume/"
      }
    }
  },
  "workspaceId": "ws-id-here"
}'

Using an Existing Kubernetes PersistentVolume

使用现有Kubernetes PersistentVolume

bash

$TFY_API_SH PUT /api/svc/v1/apps '{
  "manifest": {
    "type": "volume",
    "name": "my-existing-vol",
    "config": {
      "type": "static",
      "persistent_volume_name": "pv-name-in-k8s"
    }
  },
  "workspaceId": "ws-id-here"
}'

bash

$TFY_API_SH PUT /api/svc/v1/apps '{
  "manifest": {
    "type": "volume",
    "name": "my-existing-vol",
    "config": {
      "type": "static",
      "persistent_volume_name": "pv-name-in-k8s"
    }
  },
  "workspaceId": "ws-id-here"
}'

Listing Volumes

查询卷列表

Via Tool Call

通过工具调用查询

tfy_applications_list(filters={"workspace_fqn": "my-cluster:my-workspace", "application_type": "volume"})

tfy_applications_list(filters={"workspace_fqn": "my-cluster:my-workspace", "application_type": "volume"})

Via Direct API

通过直接API查询

bash

undefined

bash

undefined

List volumes in a workspace

查询工作空间下的卷列表

$TFY_API_SH GET '/api/svc/v1/apps?workspaceFqn=my-cluster:my-workspace&applicationType=volume'

Get a specific volume by ID

通过ID查询指定卷

$TFY_API_SH GET /api/svc/v1/apps/VOLUME_APP_ID

undefined

$TFY_API_SH GET /api/svc/v1/apps/VOLUME_APP_ID

undefined

Presenting Volumes

卷列表展示格式

Volumes in my-cluster:my-workspace:
| Name           | Size   | Storage Class | Status   | Created            |
|----------------|--------|---------------|----------|--------------------|
| training-data  | 100Gi  | efs-sc        | RUNNING  | 2026-02-10 14:30   |
| model-cache    | 50Gi   | premium-rwx   | RUNNING  | 2026-02-08 09:15   |

my-cluster:my-workspace下的卷：
| 名称           | 容量   | 存储类 | 状态   | 创建时间            |
|----------------|--------|---------------|----------|--------------------|
| training-data  | 100Gi  | efs-sc        | RUNNING  | 2026-02-10 14:30   |
| model-cache    | 50Gi   | premium-rwx   | RUNNING  | 2026-02-08 09:15   |

Attaching Volumes to Services and Jobs

将卷挂载到服务和任务

Volumes are mounted into containers at a specified path. The volume must be in the same workspace as the application.

卷会被挂载到容器的指定路径。卷必须和应用属于同一工作空间。

SDK (in deploy.py)

SDK（deploy.py中使用）

python

from truefoundry.deploy import Service, VolumeMount

service = Service(
    name="my-service",
    # ... image, ports, resources ...
    mounts=[
        VolumeMount(
            mount_path="/data",
            volume_fqn="my-cluster:my-workspace:my-volume",
        ),
    ],
)

python

from truefoundry.deploy import Service, VolumeMount

service = Service(
    name="my-service",
    # ... 镜像、端口、资源配置 ...
    mounts=[
        VolumeMount(
            mount_path="/data",
            volume_fqn="my-cluster:my-workspace:my-volume",
        ),
    ],
)

API Manifest (Service)

API配置清单（服务）

json

{
  "manifest": {
    "kind": "Service",
    "name": "my-service",
    "image": {"type": "image", "image_uri": "my-image:latest"},
    "mounts": [
      {
        "type": "volume",
        "mount_path": "/data",
        "volume_fqn": "my-cluster:my-workspace:my-volume"
      }
    ],
    "resources": {
      "cpu_request": 0.5, "cpu_limit": 1.0,
      "memory_request": 512, "memory_limit": 1024
    }
  },
  "workspaceId": "ws-id-here"
}

json

{
  "manifest": {
    "kind": "Service",
    "name": "my-service",
    "image": {"type": "image", "image_uri": "my-image:latest"},
    "mounts": [
      {
        "type": "volume",
        "mount_path": "/data",
        "volume_fqn": "my-cluster:my-workspace:my-volume"
      }
    ],
    "resources": {
      "cpu_request": 0.5, "cpu_limit": 1.0,
      "memory_request": 512, "memory_limit": 1024
    }
  },
  "workspaceId": "ws-id-here"
}

API Manifest (Job)

API配置清单（任务）

json

{
  "manifest": {
    "kind": "Job",
    "name": "my-training-job",
    "image": {"type": "image", "image_uri": "my-training:latest"},
    "mounts": [
      {
        "type": "volume",
        "mount_path": "/data",
        "volume_fqn": "my-cluster:my-workspace:training-data"
      },
      {
        "type": "volume",
        "mount_path": "/checkpoints",
        "volume_fqn": "my-cluster:my-workspace:checkpoint-vol"
      }
    ],
    "resources": {
      "cpu_request": 4.0, "cpu_limit": 8.0,
      "memory_request": 16384, "memory_limit": 32768
    }
  },
  "workspaceId": "ws-id-here"
}

json

{
  "manifest": {
    "kind": "Job",
    "name": "my-training-job",
    "image": {"type": "image", "image_uri": "my-training:latest"},
    "mounts": [
      {
        "type": "volume",
        "mount_path": "/data",
        "volume_fqn": "my-cluster:my-workspace:training-data"
      },
      {
        "type": "volume",
        "mount_path": "/checkpoints",
        "volume_fqn": "my-cluster:my-workspace:checkpoint-vol"
      }
    ],
    "resources": {
      "cpu_request": 4.0, "cpu_limit": 8.0,
      "memory_request": 16384, "memory_limit": 32768
    }
  },
  "workspaceId": "ws-id-here"
}

Volume FQN Format

卷FQN格式

The volume FQN follows the pattern:

{cluster}:{workspace}:{volume-name}

Example:

my-cluster:my-workspace:training-data

卷FQN遵循以下模式：

{集群}:{工作空间}:{卷名称}

示例：

my-cluster:my-workspace:training-data

LLM Cache Volumes

LLM缓存卷

For LLM deployments, TrueFoundry supports a

cache_volume

shorthand that creates a volume for model weight caching. This avoids re-downloading large models on every pod restart. See the

llm-deploy

skill for details.

yaml

undefined

针对LLM部署，TrueFoundry支持

cache_volume

简写配置，可创建用于模型权重缓存的卷，避免每次Pod重启都重新下载大模型。详情请查看

llm-deploy

skill。

yaml

undefined

In LLM deployment manifest

LLM部署配置清单中

cache_volume: cache_size: 50 storage_class: efs-sc

undefined

cache_volume: cache_size: 50 storage_class: efs-sc

undefined

Volume Sizing Guidelines

卷容量规划指南

Use Case	Recommended Size	Notes
Small model cache (< 7B params)	20-50 Gi	2x the model size in FP16
Large model cache (7B-70B params)	50-200 Gi	2x the model size; account for multiple formats
Shared training dataset	50-500 Gi	Depends on dataset size; leave 20% headroom
Checkpointing	20-100 Gi	Depends on checkpoint frequency and model size
General shared storage	10-50 Gi	Start small, expand as needed

Sizing tips:

Always add 20% headroom above your expected data size
Volume size can be expanded later but never reduced -- start conservatively if unsure
For model caching, use 2x the model's disk size to account for download + extraction
Monitor volume usage after deployment and expand proactively before hitting limits

场景	推荐容量	说明
小模型缓存（<7B参数）	20-50 Gi	FP16精度下模型大小的2倍
大模型缓存（7B-70B参数）	50-200 Gi	模型大小的2倍；预留多格式存储空间
共享训练数据集	50-500 Gi	取决于数据集大小；预留20%冗余空间
检查点存储	20-100 Gi	取决于检查点频率和模型大小
通用共享存储	10-50 Gi	从小容量开始，按需扩容

容量规划提示：

始终在预期数据大小基础上增加20%冗余空间
卷容量后续可扩容但永远无法缩容 -- 如果不确定请保守配置初始容量
模型缓存场景请配置模型磁盘大小的2倍容量，预留下载+解压空间
部署后监控卷使用情况，在达到容量上限前主动扩容

Static Volume Setup

静态卷设置

For detailed setup instructions for AWS EFS, AWS S3, GCP GCS Fuse, and Azure Files/Blob, see

references/static-volume-setup.md

AWS EFS、AWS S3、GCP GCS Fuse和Azure Files/Blob的详细设置说明请查看

references/static-volume-setup.md

。

Volume Browser

卷浏览器

For Volume Browser configuration fields, setup steps, and access instructions, see

references/volume-browser-setup.md

</instructions>

<success_criteria>

The agent asked "create new or use existing?" before proceeding
The agent has confirmed volume name, size, storage class, and workspace with the user before creating
The agent asked whether to enable Volume Browser and collected endpoint/password details if yes
The volume was successfully created and is in RUNNING status
The user can list all volumes in their target workspace
The user can attach the volume to a service or job using the correct volume FQN
The agent has advised on appropriate sizing based on the user's use case
The user understands the difference between volumes and blob storage for their scenario

</success_criteria>

卷浏览器的配置字段、设置步骤和访问说明请查看

references/volume-browser-setup.md

。

</instructions>

<success_criteria>

Agent在执行操作前询问用户「是要创建新卷还是使用现有卷？」
Agent在创建前已和用户确认卷名称、容量、存储类和工作空间
Agent询问用户是否启用卷浏览器，若启用则收集访问地址/密码相关信息
卷创建成功且处于RUNNING状态
用户可以查询目标工作空间下的所有卷
用户可以使用正确的卷FQN将卷挂载到服务或任务
Agent根据用户场景提供了合适的容量建议
用户理解卷和对象存储在自身场景下的差异

</success_criteria>

Composability

组合使用说明

Before deploying with volumes: Use
```
workspaces
```
skill to get workspace FQN, then create the volume in the same workspace
With secrets skill: Create a password secret before enabling Volume Browser (password_secret_fqn is required)
With deploy skill: After creating a volume, add
```
VolumeMount
```
to the service's deploy.py to attach it
With llm-deploy skill: Use
```
cache_volume
```
in LLM deployment manifests for model weight caching
With jobs skill: Mount volumes to training jobs for checkpointing and shared data access
With applications skill: List volumes alongside other application types to see what storage exists
After creating: Use
```
applications
```
skill to verify the volume was created successfully

</references> <troubleshooting>

使用卷部署前：使用
```
workspaces
```
skill获取工作空间FQN，然后在同一工作空间下创建卷
结合secrets skill使用：启用卷浏览器前先创建密码密钥（必填项password_secret_fqn）
结合deploy skill使用：创建卷后，在服务的deploy.py中添加
```
VolumeMount
```
完成挂载
结合llm-deploy skill使用：在LLM部署配置中使用
```
cache_volume
```
实现模型权重缓存
结合jobs skill使用：将卷挂载到训练任务，用于存储检查点和共享数据
结合applications skill使用：将卷和其他应用类型一同查询，了解现有存储资源
创建完成后：使用
```
applications
```
skill验证卷是否创建成功

</references> <troubleshooting>

Error Handling

错误处理

Error	Cause	Fix
Volume not found	Wrong name or workspace	Verify FQN; volumes are workspace-scoped
Storage class not available	Cluster missing provisioner	Check `GET /api/svc/v1/clusters/CLUSTER_ID` for available classes
Size cannot be reduced	PVC limitation	Create new smaller volume and migrate data
Workspace mismatch	Volume in different workspace	Create volume in same workspace as the app
Permission denied	API key lacks access	Check API key permissions for this workspace
PV not found (static)	K8s PV doesn't exist	Verify with `kubectl get pv <pv-name>`
Data corruption	Multiple pods writing same path	Use per-pod sub-directories (e.g., `/data/pod-{POD_NAME}/` )

</troubleshooting>

错误	原因	解决方案
卷未找到	名称或工作空间错误	验证FQN；卷是工作空间级别的资源
存储类不可用	集群缺少对应配置器	调用 `GET /api/svc/v1/clusters/CLUSTER_ID` 查看可用存储类
无法缩容	PVC固有约束	创建新的小容量卷并迁移数据
工作空间不匹配	卷和应用属于不同工作空间	在应用所属的同一工作空间下创建卷
权限不足	API密钥缺少访问权限	检查API密钥对当前工作空间的权限
PV未找到（静态卷场景）	K8s PV不存在	执行 `kubectl get pv <pv-name>` 验证
数据损坏	多个Pod同时写入同一路径	使用Pod专属子目录（例如 `/data/pod-{POD_NAME}/` ）

</troubleshooting>