name: alibabacloud-pai-eas-service-deploy description: | Deploy AI models as PAI-EAS inference services. Supports LLMs (Qwen, Llama), image gen (SD, SDXL), speech synthesis, and more. When to use: deploy models, create inference services, EAS deployment, model serving, deploy vLLM/SGLang/ComfyUI. license: Apache-2.0 metadata: version: "1.0.0" domain: aiops owner: pai-eas-team contact: pai-eas-agent@alibaba-inc.com tags: - pai-eas - model-deployment - inference-service - llm - vllm - sglang required_tools: - aliyun - jq prerequisites: - "Aliyun CLI >= 3.3.3" - "jq command-line JSON processor" required_permissions: - "eas:CreateService" - "eas:DescribeService" - "eas:ListServices" - "eas:DescribeMachineSpec" - "eas:ListResources" - "eas:ListGateway" - "eas:DescribeGateway" - "nlb:ListLoadBalancers" - "aiworkspace:ListImages" - "aiworkspace:ListWorkspaces" - "vpc:DescribeVpcs" - "vpc:DescribeVSwitches" - "ecs:DescribeSecurityGroups"

PAI-EAS Service Deployment

⚠️ TOP RULES (read first)

1. 🔴 NO DUPLICATE SERVICE NAMES 🔴

If a service with the target name already exists: STOP and inform the user. Do NOT delete and recreate. Do NOT reuse it either.

2. Mandatory API Calls — Execute ALL of these in order:

Execute #1 and #2 ALWAYS, even if user provided the info. describe-machine-spec ≠ list-resources. describe-service ≠ ListServices.

3. Prohibited — ❌ Reuse existing service ❌ Write bash scripts (run CLI directly) ❌ CPU+vLLM/SGLang ❌ file:// in create-service ❌ Skip mandatory APIs ❌ Change the service name the user specified ❌ Poll describe-service in a loop (call once only) ❌ Stop after writing/validating service.json — writing the JSON is NOT completing the task. You MUST run create-service (Step 6) then describe-service (Step 7). If resources were found, deploy them.

4. Autonomous Execution — Do NOT ask user for info discoverable via APIs. Do NOT ask "should I proceed?" Execute directly. Timeout? Retry with --read-timeout 60. Error? Inform user and CONTINUE. Missing param? Pick reasonable default. If any pre-check or resource discovery step fails, log the failure and continue to the next step. Only STOP for the specific conditions listed in Self-Verify Checkpoints (duplicate service name, missing NLB/GW/dedicated resource group).

5. Self-Verify Checkpoints:

Before Step 2: Did Step 1.5 confirm no duplicate service name?
  If duplicate → STOP, inform user, do NOT proceed.
Before Step 5: Have I run list-images AND describe-machine-spec?
  If NO → STOP and run them NOW.
Before Step 6: Did Step 1.5 confirm no duplicate service name?
  If duplicate exists → STOP, inform user, do NOT proceed.
Step 4 resources: If NLB/GW/Resource Group not found → inform user and STOP.
  Do NOT block or attempt workarounds.
After Step 7: Did I call describe-service once and report the status?

6. Run CLI commands DIRECTLY — Use execute_shell_command to run each aliyun CLI command one at a time. Do NOT write bash scripts, deployment scripts, or shell files. Do NOT use retry_command() or any wrapper functions. Each command = one direct execution.

Pre-checks

# Install or upgrade Aliyun CLI to >= 3.3.3 (required baseline), then verify.
curl -fsSL https://aliyuncli.alicdn.com/install.sh | bash
aliyun version   # must report >= 3.3.3; re-run install.sh to upgrade if lower

# Generate the per-session trace id once (see Observability section) and build
# the unified user-agent. Pass it via --user-agent on every CLI command.
SESSION_ID=$(openssl rand -hex 16)
UA="AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy/${SESSION_ID}"
aliyun plugin update
aliyun configure set --auto-plugin-install true

⚠️ MANDATORY: Install and update required plugins BEFORE proceeding:

aliyun plugin update
aliyun plugin install aiworkspace --force
aliyun plugin install eas --force

Verify plugins are installed:

aliyun plugin list 2>/dev/null | grep -E "aiworkspace|eas"

If either plugin is missing after install, retry with:

aliyun plugin install <plugin-name> --force

See CLI Guide. Never read/echo AK/SK. Timeout? Retry with --read-timeout 60.

⚠️ MANDATORY CHECKPOINT: After pre-checks, you MUST proceed to Step 1 regardless of any warnings during plugin installation. Plugin install failures are non-blocking — the plugins may already be installed. NEVER stop execution at the pre-check stage.

Observability

All CLI, SDK, and Terraform calls MUST carry a unified user-agent (UA) so requests trace back to this skill and one session.

UA template declaration (placeholder form):

--user-agent AlibabaCloud-Agent-Skills/{SKILL_NAME}/{session-id}

• {SKILL_NAME} — fixed identifier: alibabacloud-pai-eas-service-deploy
• {session-id} — 32-char lowercase hex (128-bit), generated once per session

session-id rule: Generate it once before any call via SESSION_ID=$(openssl rand -hex 16); reuse the same value for every CLI (--user-agent), SDK (client UA), and Terraform (ALIBABACLOUDSTACK_UA / provider ua) call so all requests share one trace id. Do NOT regenerate between steps or hardcode it.

SESSION_ID=$(openssl rand -hex 16)
UA="AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy/${SESSION_ID}"
# Pass on every command, e.g.:
aliyun eas describe-service --cluster-id <region> --service-name <service> --user-agent "$UA"

Step 1: Collect Parameters

Service name: lowercase/digits/underscores only. No hyphens. 3-63 chars. IMPORTANT: Use the EXACT service name the user specifies. Do NOT rename. If the user specifies a prefix (e.g. "starts with skill_qwen_"), generate a random suffix of 6 digits (e.g. skill_qwen_482917).

Set profile region — Set the CLI profile region to match the deployment region. This avoids "Region mismatch" errors when --cluster-id differs from the profile's default region:

aliyun configure set --region <region>

Workspace ID: Required in metadata.workspace_id. If user does not specify a workspace, query available workspaces and pick one:

aliyun aiworkspace list-workspaces --region <region> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '.Workspaces[] | select(.Status == "ENABLED") | {WorkspaceId, WorkspaceName}'

If multiple workspaces exist, list them and let the user choose. If only one exists, use it directly.

Step 1.5: Check for Duplicate Service Name

aliyun eas list-services --region <region> --cluster-id <region> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '.ServiceList[] | select(.ServiceName == "<name>") | {ServiceName, Status}'

If a service with the same name already exists → STOP and inform the user: "A service named already exists (Status: ). Please choose a different name." Do NOT delete or reuse it.

If no duplicate → proceed to Step 2.

Step 2: ListImages (🚧 BLOCKING GATE — NEVER SKIP)

Execute even if user provided image URI. Purpose = VALIDATION.

⚠️ If you see "parse error" or "Exit Code 4", the plugin failed to install. You MUST retry with explicit install:

aliyun plugin install aiworkspace --force

Then retry the list-images command. Do NOT skip this step.

aliyun aiworkspace list-images --region <region> --verbose true \
  --labels 'system.official=true,system.supported.eas=true' \
  --page-size 50 --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '.Images[] | {ImageId, ImageUri, ImageCategory}' > /tmp/eas_images.json
echo "Saved $(jq length /tmp/eas_images.json) images"

If still fails after plugin reinstall, retry with timeout:

aliyun aiworkspace list-images --region <region> --verbose true \
  --labels 'system.official=true,system.supported.eas=true' \
  --page-size 50 --read-timeout 60 \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '.Images[] | {ImageId, ImageUri, ImageCategory}' > /tmp/eas_images.json
echo "Saved $(jq length /tmp/eas_images.json) images"

See Model-Image. This output is REQUIRED by Step 5. Do NOT skip. ❌ NEVER skip this step saying "user already provided the image".

Step 3: describe-machine-spec (🚧 BLOCKING GATE — NEVER SKIP)

Execute even if user specified instance type. NOT list-resources.

⚠️ If this command fails, retry with --read-timeout 60. NEVER proceed to Step 5 without /tmp/eas_specs.json existing.

aliyun eas describe-machine-spec --region <region> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '.InstanceTypes[] | select(.GPU > 0) | {InstanceType, CPU, GPU, Memory}' \
  > /tmp/eas_specs.json
echo "Available GPU specs:"; cat /tmp/eas_specs.json

CPU→GPU auto-switch: If vLLM/SGLang + CPU, silently pick GPU. Do NOT ask. This output is REQUIRED by Step 5. Do NOT skip.

Step 4: Network & Resource Config

⚠️ If a required resource does not exist → STOP and inform the user. Do NOT block or attempt workarounds. This is a valid outcome.

Dedicated Gateway — Call list-gateway. If no gateway exists → inform user and STOP. Otherwise call describe-gateway to get VPC/VSwitch, then query security group under that VPC. If no security group found → inform user and STOP.

aliyun eas list-gateway --region <region> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy

If gateway found, get details:

aliyun eas describe-gateway --region <region> --cluster-id <region> \
  --gateway-id <gateway_id> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy

Extract VPC and comma-separated VSwitch ID:

aliyun eas describe-gateway --region <region> --cluster-id <region> \
  --gateway-id <gateway_id> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '{vpc_id: .LoadBalancerList[0].VpcId, vswitch_id: (.LoadBalancerList[0].VSwitchIds | join(","))}'

NLB — Requires VPC/VSwitch/SecurityGroup. If user does not provide them, query via APIs. If any required resource not found → inform user and STOP. ⚠️ NLB requires ≥2 VSwitches across different availability zones. Use comma-separated format: "vswitch_id": "vsw-zone-a,vsw-zone-b". ⚠️ NLB Plugin Bug (aliyun-cli-eas v0.2.0): If create-service with NLB config returns 400 with 'vswitch can not be null' or 'vpcId, vswId and securityGroupId are required', this is a known CLI plugin bug (not a resource issue). Fallback strategy: 1. Retry create-service with NLB config once more (max 2 attempts). 2. If both fail → Remove networking.nlb and cloud.networking from service.json, redeploy with shared gateway. 3. Inform user: "NLB config failed due to CLI plugin limitation. Deployed with shared gateway instead."

EAS Dedicated Resource Group — Call list-resources. Filter for ResourceType == "Dedicated" and Status == "ResourceReady".

aliyun eas list-resources --region <region> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '.Resources[] | select(.ResourceType == "Dedicated" and .Status == "ResourceReady") | {ResourceId, ResourceType, Status}'

• If exists → Set "metadata": {"resource": "<ResourceId>"}. Do NOT set cloud.computing.
• If NOT exists → Inform the user and STOP. Do NOT fall back to public resource group.

Step 5: Build Service JSON

⚠️ BEFORE building JSON, you MUST read these reference files: - references/config-patterns.md — Complete JSON templates for all 8 patterns - references/config-schema.md — Field descriptions and validation rules - references/storage-mount.md — OSS/NAS mount configuration details - references/network-config.md — NLB/Gateway network configuration details

⚠️ HARD GATE: Before writing service.json, VERIFY these files exist and have content. If either is missing → STOP and run that Step NOW.

test -s /tmp/eas_images.json || echo "MISSING: Run Step 2 NOW"
test -s /tmp/eas_specs.json || echo "MISSING: Run Step 3 NOW"

⚠️ JSON format rules: - Allowed top-level keys: metadata, containers, storage, cloud, autoscaler, networking - ❌ NEVER use as top-level keys: spec, ServiceName, Image, Cpu, Memory, Gpu, processor_path, resourceGroupId, instance, port, command, access - ❌ FORBIDDEN fields: processor_path, resourceGroupId, spec, access - metadata.name = service name, metadata.workspace_id = workspace (REQUIRED) - containers[].image = image URI, containers[].command = start command, containers[].port = port - cloud.computing.instance_type = instance type (MANDATORY for shared gateway)

Quick Reference — JSON Skeletons

Below are minimal skeletons. Read references/config-patterns.md for complete templates with all fields and examples.

Base (Shared Gateway):

{"metadata":{"name":"<name>","instance":1,"workspace_id":"<ws>"},
 "containers":[{"image":"<img>","port":<p>,"command":"<cmd>"}],
 "cloud":{"computing":{"instance_type":"<type>"}}}

+ OSS → add "storage":[{"mount_path":"/dir","oss":{"path":"oss://<b>/<p>/","readOnly":true}}] + Autoscaling → add "autoscaler":{"min":1,"max":4,"scaleStrategies":[{"metricName":"qps","threshold":20}]} + Health Check → add startup_check to containers[] (see config-patterns.md Pattern 4)

NLB — full template (read references/network-config.md for details):

{"metadata":{"name":"<name>","instance":1,"workspace_id":"<ws>"},
 "containers":[{"image":"<img>","port":<p>,"command":"<cmd>"}],
 "cloud":{"computing":{"instance_type":"<type>"},
          "networking":{"vpc_id":"<vpc>","vswitch_id":"<vsw1>,<vsw2>","security_group_id":"<sg>"}},
 "networking":{"nlb":[{"id":"default","listener_port":<p>,"netType":"intranet"}]}}

⚠️ vswitch_id must be comma-separated with ≥2 VSwitches across different zones

Dedicated Resource Group — "metadata.resource" instead of cloud.computing:

{"metadata":{"name":"<name>","instance":1,"resource":"<res_id>","workspace_id":"<ws>"},
 "containers":[{"image":"<img>","port":<p>,"command":"<cmd>"}]}

Dedicated Gateway — networking.gateway + cloud.networking:

{"metadata":{"name":"<name>","instance":1,"workspace_id":"<ws>"},
 "containers":[{"image":"<img>","port":<p>,"command":"<cmd>"}],
 "networking":{"gateway":"<gw_id>"},
 "cloud":{"computing":{"instance_type":"<type>"},
          "networking":{"vpc_id":"<vpc>","vswitch_id":"<vsw1>,<vsw2>","security_group_id":"<sg>"}}}

⚠️ vswitch_id comma-separated if gateway returns multiple VSwitches

Validate Before Writing

jq -r '.[] | select(.ImageUri | contains("vllm")) | .ImageUri' /tmp/eas_images.json
jq -r '.[] | select(.InstanceType == "<type>") | .InstanceType' /tmp/eas_specs.json

➡️ After service.json validates, IMMEDIATELY go to Step 6 — a validated JSON is not a deployment.

Step 6: Create Service (MANDATORY)

🔴 CONFIRM: Did Step 1.5 confirm no duplicate service name? If a service with this name already exists → STOP. Inform the user and do NOT proceed with create-service. Use $(cat service.json) NOT file://service.json. Run this DIRECTLY via execute_shell_command, do NOT write a bash script.

aliyun eas create-service --region <region> \
  --body "$(cat service.json)" \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy

409 Conflict → Service already exists. Inform the user and STOP. 400 BadRequest with 'vswitch can not be null' or 'vpcId, vswId and securityGroupId are required' → NLB CLI plugin bug (see Step 4 fallback). Remove networking.nlb and cloud.networking from service.json and retry.

Step 7: Verify Deployment

Call describe-service ONCE to check the current status. Do NOT poll. Do NOT loop. Do NOT wait for Running.

aliyun eas describe-service --region <region> --cluster-id <region> \
  --service-name <name> \
  --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '{Status, ServiceName, ServiceId}'

Report whatever status you get (Running, Waiting, Creating, etc.) and proceed to Step 8 immediately. create-service returning 200 = success.

Step 8: Report Result (MANDATORY)

Get endpoint info via DescribeServiceEndpoint:

aliyun eas describe-service-endpoints --region <region> --cluster-id <region> \
  --service-name <name> --user-agent AlibabaCloud-Agent-Skills/alibabacloud-pai-eas-service-deploy | \
  jq '{AccessToken, Endpoints: [.Endpoints[] | {
    Type: .EndpointType, Port: .Port,
    InternetEndpoints: .InternetEndpoints,
    IntranetEndpoints: .IntranetEndpoints
  }]}'

Use the status from Step 7 and the endpoints above to report.

Copy the ENTIRE output into your final response. Format:

Deployment Summary
==================
Service Name: <name>
Status: <from Step 7>

Endpoints:
- <EndpointType>:
    InternetEndpoint: <url or null>
    IntranetEndpoint: <url or null>
    Port: <port or 0>

Service Invocation Examples:
  curl <internet-endpoint>/api/predict/<name> \
    -H "Authorization: <AccessToken>"
  curl <intranet-endpoint>/api/predict/<name> \
    -H "Authorization: <AccessToken>"
  curl <nlb-domain>:<listener_port>/api/predict/<name> \
    -H "Authorization: <AccessToken>"

InternetEndpoint and IntranetEndpoint MUST appear in your response, even if null. If null: (not available for this network type)

Always include a service invocation example using the AccessToken and endpoint URL.

Success criteria: create-service returning 200 with ServiceId = success. Any status (Running, Waiting, Creating) is acceptable.

References (read when needed)

这个 Skill 质量很好，文档非常详细，连操作步骤和风险提示都写得清清楚楚，用起来不容易出错。不过有些地方小有不足，比如许可证说明前后不太一致，工具本身也有点小问题（NLB 配置有时候会失败）。整体来说非常实用，值得推荐。