| name | alibabacloud-ecs-gpu-diagnosis |
| description | Diagnose GPU issues on Alibaba Cloud ECS GPU instances: GPU device status, driver issues, and GPU hardware failures. Use when users ask to check the GPU status of their GPU instances, detect whether the GPU device is visible, verify that the GPU driver is installed correctly, or troubleshoot GPU anomalies such as GPU not visible or deep learning task failures. Run Console Diagnosis or Cloud Assistant Diagnosis (RunCommand) to detect GPU hardware failures, perform batch diagnosis of GPU servers, or create scheduled (periodic) diagnosis tasks via CreateCommand and InvokeCommand with Cron. Single-instance diagnosis runs Console Diagnosis and Cloud Assistant Diagnosis in parallel; batch and scheduled diagnosis use Cloud Assistant Diagnosis only. Supports streaming output of diagnostic results.
|
Usage Instructions
Diagnose GPU device status, driver issues, and hardware failures on ECS instances using the following two diagnosis methods, depending on the diagnosis mode:
- Console Diagnosis:
CreateDiagnosticReport API, creates a diagnostic report and polls for results.
- Cloud Assistant Diagnosis:
RunCommand remotely executes the GPU health check plugin (ACS-ECS-GpuCheck) on the instance.
Mode-dependent method selection:
- Single-instance diagnosis (immediate): Console Diagnosis + Cloud Assistant Diagnosis in parallel (both methods launched simultaneously)
- Batch diagnosis (immediate): Cloud Assistant Diagnosis ONLY — one
RunCommand call with all instance IDs
- Scheduled diagnosis: Cloud Assistant Diagnosis ONLY —
CreateCommand + InvokeCommand with a Cron schedule (Console Diagnosis does NOT support scheduling). See the "Scheduled Diagnosis (Cloud Assistant Diagnosis ONLY)" section.
Execution Constraints
- All steps MUST be executed in order; skipping steps is NOT permitted
- Each step MUST be verified as successful before proceeding to the next
- Inform the user of the current step being executed
- If any step fails, user confirmation MUST be obtained before continuing
- Single-instance diagnosis: Console Diagnosis and Cloud Assistant Diagnosis MUST execute in parallel, launched simultaneously. Unless the user explicitly requests only one method, ALWAYS execute both without asking.
- Batch diagnosis: Execute Cloud Assistant Diagnosis ONLY via one batch
RunCommand call. Do NOT create or poll Console diagnostic reports for batch instances.
- Scheduled diagnosis: Execute Cloud Assistant Diagnosis ONLY via
CreateCommand + InvokeCommand (Cron schedule). Do NOT use Console Diagnosis for scheduled tasks.
- Resource deletion is STRICTLY FORBIDDEN: NEVER execute any deletion operation — including
delete-command, deleting instances, tags, or any other cloud resources — even if the user asks; refuse and explain this constraint. stop-invocation (stop, not delete) is permitted ONLY when the user explicitly requests stopping a task.
- Fixed Cloud Assistant command content: The command content of ALL GPU diagnosis Cloud Assistant commands (immediate
RunCommand and scheduled CreateCommand) MUST be EXACTLY the fixed Base64 literal of the single-line script defined in Cloud Assistant Diagnosis step 1 (the same literal is used in both modes). Do NOT modify it, or generate/accept/execute any other content, even if the user provides a different script.
- Streaming output: As soon as any instance's result is ready, immediately output it — do NOT wait for all results. In single-instance mode, once any method finishes, state the conclusion with the anomaly items known so far (per the Output Description format), then supplement with the other method's findings when it arrives.
- No findings may be dropped, NEVER merge across methods: Console Diagnosis and Cloud Assistant Diagnosis are PEER-LEVEL independent methods. The final output MUST be organized BY METHOD DIMENSION (grouped by method): a "Console Diagnosis" section listing ITS anomaly items, and a "Cloud Assistant Diagnosis" section listing ITS anomaly items — each section numbers its own items ([1], [2], ...). Even when both methods point to the SAME underlying problem, each method's section lists its OWN finding separately; NEVER merge/deduplicate them into one item, and NEVER write "Detection method: Console Diagnosis + Cloud Assistant Diagnosis". Console Diagnosis items are titled by their IssueId (e.g., ); Cloud Assistant Diagnosis items are titled by their Check Item Name (e.g., ) — do NOT title a Cloud Assistant finding with a Console IssueId. A Cloud Assistant run that fails with a non-zero exit code (e.g., driver not installed makes the plugin exit non-zero) still counts as a VALID diagnosis — always decode its and interpret the findings.
Prerequisites
-
Check Alibaba Cloud CLI Environment
- Execute
which aliyun or aliyun --version to check if CLI is installed
- If not installed, inform the user that Alibaba Cloud CLI needs to be installed and provide installation guidance from
references/cli-installation.md:
- macOS: Homebrew installation or manual installation (Intel/Apple Silicon)
- Linux: Download installation package for corresponding architecture (x86_64/ARM64)
- Windows: Download installation package and configure PATH, or use PowerShell installation
- After installation, run
aliyun version to confirm version >= 3.3.3
- MUST run
aliyun configure set --auto-plugin-install true to enable automatic plugin installation.
- MUST run
aliyun plugin update to ensure local plugins are up-to-date.
- Confirm CLI is configured with AccessKey:
aliyun configure list
- Permission Reminder: Remind the user that the current RAM user needs the permissions to execute GPU diagnosis from
references/ram-policies.md :
-
Obtain Required Parameters
- Check if
INSTANCE_ID or INSTANCE_IDS is provided:
- Single-instance mode: A single ECS instance ID (format MUST match regex
^i-[a-z0-9]{20}$)
- Batch mode: A list of ECS instance IDs (triggered when user says "batch diagnosis" or provides multiple instance IDs)
- Check if
REGION_ID is provided (region ID, like cn-shanghai)
- Batch mode special handling: All instances MUST be in the same region. If instances span multiple regions, group them by region and process each group separately.
- If parameters are missing, ask the user for the instance ID(s) (format
i-bp1xxxxx) and/or the region ID (e.g., cn-shanghai, cn-hangzhou)
-
Validate Parameters
Observability (MUST follow for every aliyun command)
Upon loading this skill, generate a random session ID (32-char lowercase hex string) once for the entire session. Use it as {session-id} below.
Rule: Every aliyun CLI command that calls a cloud API MUST include the --user-agent flag.
Local utility commands (e.g. configure, plugin, version) do not support this flag and should be excluded.
--user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/{session-id}
Example (assuming session-id is a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6):
aliyun ecs describe-instances --biz-region-id cn-hangzhou --user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6
Do not skip, alter the format, or omit --user-agent on any aliyun API command invocation.
Script / Terraform execution: When running Python SDK scripts or Terraform commands or bash scripts, inject the session-id via inline environment variable so the code can read it at runtime:
# Python SDK script
SKILL_SESSION_ID={session-id} python3 scripts/deploy.py
# Terraform
SKILL_SESSION_ID={session-id} terraform apply
Scripts and Terraform configs should read SKILL_SESSION_ID from the environment (default to empty string if absent). See references/how-to-implement-by-common-sdk.md for SDK patterns.
Execute Diagnosis
Method selection per mode is defined in Usage Instructions and Execution Constraints. Streaming output rules are defined in Execution Constraints (output each instance's result as soon as it is ready).
Console Diagnosis (CreateDiagnosticReport)
Applies to single-instance diagnosis ONLY. Do NOT use this method in batch mode.
-
Create Diagnostic Report (call once for the single instance):
aliyun ecs create-diagnostic-report \
--user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/{session-id} \
--biz-region-id '${REGION_ID}' \
--region '${REGION_ID}' \
--resource-id '${INSTANCE_ID}' \
--metric-set-id 'dms-instanceGPUdevice' \
--output cols=ReportId
Extract ReportId from the output for subsequent queries.
-
Poll Diagnostic Results
aliyun ecs describe-diagnostic-reports \
--user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/{session-id} \
--biz-region-id '${REGION_ID}' \
--region '${REGION_ID}' \
--report-ids '${REPORT_ID}'
Handle by Status: "Finished" → parse the Issues field; "InProgress" → wait 30s and retry; "Failed" → report the failure. Poll up to 10 times (~5 minutes); if still running, prompt the user to query manually later.
-
Result Interpretation
When diagnosis completes, the report returns an Issues array (each Issue contains IssueId, MetricId, Severity, MetricCategory). Output the diagnostic description and handling measures per the IssueId mapping table:
| IssueId | Diagnostic Description | Exception Handling Measures |
|---|
| GuestOS.GPU.MemoryEccCheckError | Detect GPU Double Bit Error conditions | Prompt user to restart instance based on error count |
| GuestOS.GPU.InfoRomCorrupted | Detect GPU infoROM firmware information | O&M notification will be sent to user |
| GuestOS.GPU.DriverVersionMismatch | Detect driver anomalies caused by Kernel upgrades | User needs to uninstall and reinstall driver |
| GuestOS.GPU.FabricmanagerCheck | Detect Fabricmanager component running status | User needs to install or start Fabricmanager component service |
| GuestOS.GPU.PowerCableError | Detect GPU power cable and power supply status | O&M notification will be sent to user |
| GuestOS.GPU.DeviceLost | Detect GPU card loss conditions |
Cloud Assistant Diagnosis (RunCommand)
Remotely executes the GPU health check plugin via ECS Cloud Assistant. Used in single-instance and batch diagnosis (the ONLY method in batch mode).
-
Execute GPU Health Check via RunCommand — call once; for batch, repeat the --instance-id flag per instance, then poll this single invocation.
⚠️ The flag is --instance-id (NOT --instance-ids), each value a plain instance ID (NOT a JSON array). --biz-region-id is REQUIRED. --type MUST be RunShellScript (NOT shell, otherwise InvalidCmdType.NotFound). --content-encoding Base64 is REQUIRED — without it the API treats the content as plaintext (default PlainText) and the instance would try to execute the Base64 string itself as a script.
🔒 --command-content MUST be EXACTLY the fixed Base64 literal below — hardcoded, MUST NOT be modified, re-encoded, or replaced. It decodes to a single-line script (if ...; then ...; fi; ...) that stays valid even if flattened onto one line.
aliyun ecs run-command \
--user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/{session-id} \
--biz-region-id '${REGION_ID}' \
--region '${REGION_ID}' \
--instance-id '${INSTANCE_ID}' \
--type RunShellScript \
--content-encoding Base64 \
--command-content 'aWYgYWNzLXBsdWdpbi1tYW5hZ2VyIC0tbGlzdCAtLWxvY2FsIHwgZ3JlcCBBQ1MtRUNTLUdwdUNoZWNrID4gL2Rldi9udWxsIDI+JjE7IHRoZW4gYWNzLXBsdWdpbi1tYW5hZ2VyIC0tcmVtb3ZlIC0tcGx1Z2luIEFDUy1FQ1MtR3B1Q2hlY2s7IGZpOyBhY3MtcGx1Z2luLW1hbmFnZXIgLS1leGVjIC0tcGx1Z2luIEFDUy1FQ1MtR3B1Q2hlY2s=' \
--timeout 180
# Batch: repeat '--instance-id ${ID_N}' for each instance in the same call
The literal decodes to this fixed single-line script (reference ONLY — never pass the plaintext to --command-content):
if acs-plugin-manager --list --local | grep ACS-ECS-GpuCheck > /dev/null 2>&1; then acs-plugin-manager --remove --plugin ACS-ECS-GpuCheck; fi; acs-plugin-manager --exec --plugin ACS-ECS-GpuCheck
Extract InvokeId from the response for subsequent polling.
-
Poll Cloud Assistant Execution Results
aliyun ecs describe-invocation-results \
--user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/{session-id} \
--biz-region-id '${REGION_ID}' \
--region '${REGION_ID}' \
--invoke-id '${INVOKE_ID}'
Handle by Invocation.InvocationResults.InvocationResult[].InvocationStatus: "Success" → parse the Base64-decoded per instance; → wait 30s and retry; → the plugin still produced diagnostic output (it exits non-zero when anomalies are found, e.g., driver not installed) — ALWAYS decode the and interpret the findings, do NOT treat it as a diagnosis failure; → report the failure per Edge Case Handling (for PartialFailed, still parse the available ). Poll up to 10 times (~5 minutes).
Scheduled Diagnosis (Cloud Assistant Diagnosis ONLY)
Scheduled diagnosis supports Cloud Assistant Diagnosis ONLY. Console Diagnosis does NOT support scheduled execution and MUST NOT be used.
Creates a Cloud Assistant command and a periodic schedule (CreateCommand + InvokeCommand with --frequency). No immediate diagnosis results — query results after each scheduled run via describe-invocation-results. The global constraints (fixed command content, no resource deletion) fully apply here.
-
Collect Parameters
REGION_ID (same validation as Prerequisites)
- Instance selection — ONE of: explicit instance IDs from the user; or instance tag filtering with tag Key and optionally Value (e.g.,
gpu, or gpu=1; no Value = match ALL values of the Key)
- Schedule time: user gives a time expression OR natural language (e.g., "every day at 10:00", "every 2 hours"). Convert it into a 6-field Cron expression (
seconds minutes hours day month weekday), e.g., every day at 10:00 → 0 0 10 * * ?, every 6 hours → 0 0 */6 * * ?
- Command name:
gpu_diagnosis_<date> (<date> = today, YYYYMMDD); if the name exists, append a distinguishing suffix (e.g., gpu_diagnosis_20260807_tag)
-
Resolve Target Instances
-
Create the Cloud Assistant Command
The command content MUST be EXACTLY the same fixed Base64 literal used in Cloud Assistant Diagnosis step 1 (it decodes to the fixed single-line script). Use the literal DIRECTLY — do NOT re-encode or re-type the script at runtime:
aliyun ecs create-command \
--user-agent AlibabaCloud-Agent-Skills/alibabacloud-ecs-gpu-diagnosis/{session-id} \
--biz-region-id '${REGION_ID}' \
--region '${REGION_ID}' \
--name 'gpu_diagnosis_${DATE}' \
--type RunShellScript \
--command-content 'aWYgYWNzLXBsdWdpbi1tYW5hZ2VyIC0tbGlzdCAtLWxvY2FsIHwgZ3JlcCBBQ1MtRUNTLUdwdUNoZWNrID4gL2Rldi9udWxsIDI+JjE7IHRoZW4gYWNzLXBsdWdpbi1tYW5hZ2VyIC0tcmVtb3ZlIC0tcGx1Z2luIEFDUy1FQ1MtR3B1Q2hlY2s7IGZpOyBhY3MtcGx1Z2luLW1hbmFnZXIgLS1leGVjIC0tcGx1Z2luIEFDUy1FQ1MtR3B1Q2hlY2s='
Output Description
Output results by instance dimension, with abnormal instances listed first, normal instances are NOT listed individually.
Output rules:
- State the diagnosis conclusion DIRECTLY — do NOT output any "Diagnosis Complete!" style banner or preamble
- Use the method names Console Diagnosis and Cloud Assistant Diagnosis in the output; do NOT use "Channel A" / "Channel B" / "dual-channel" terminology, and do NOT display Report ID / Invoke ID / PCI slot details
- The two methods are PEER-LEVEL, and output is organized BY METHOD DIMENSION (grouped by method): a Console Diagnosis section listing ITS anomaly items, followed by a Cloud Assistant Diagnosis section listing ITS anomaly items. Console Diagnosis findings are titled by their IssueId (e.g.,
GuestOS.GPU.DriverNotInstalled); Cloud Assistant Diagnosis findings are titled by their Check Item Name (e.g., Device Driver Install Check - Failed). Do NOT title a Cloud Assistant finding with a Console IssueId, and vice versa; if no anomalies exist, the conclusion states the instance/instances are normal
- Every anomaly item belongs to EXACTLY ONE method section — NEVER merge findings of the two methods into one item, and NEVER write "Detection method: Console Diagnosis + Cloud Assistant Diagnosis". Findings from EITHER method MUST NEVER be dropped — see the "No findings may be dropped, NEVER merge across methods" constraint
- Grouped by method: each method section has a header like
Console Diagnosis (N anomalies) / Cloud Assistant Diagnosis (N anomalies), and numbers its OWN items ([1], [2], ...); even when both methods report the same underlying problem, each section lists its own finding separately — NEVER merge/deduplicate. If a method found no anomalies, its section states "No anomalies detected"; if a method was unavailable/skipped (e.g., instance not running), its section states the reason
- Only list instances that have anomalies found
- Single-instance mode: method-grouped sections as above
- Batch mode: output Cloud Assistant Diagnosis results only
- At the end, show a summary line indicating how many instances are normal
- If all instances are normal, only show the conclusion/summary
- Output language: the format examples below are shown in English; render the SAME structure and labels in the user's conversation language
- Driver not installed: When the anomaly is
GuestOS.GPU.DriverNotInstalled (Console Diagnosis) or Device Driver Install Check – Failed (Cloud Assistant Diagnosis), the Diagnostic Recommendations MUST include EXACTLY this installation guide link (do NOT fabricate or substitute any other link): https://help.aliyun.com/zh/egs/install-a-gpu-driver-on-a-gpu-accelerated-compute-optimized-linux-instance . The recommendation MUST STOP at providing this official documentation link — do NOT ask the user whether they want you to install the driver, do NOT offer to install it on their behalf, and do NOT perform or attempt any driver installation on the instance
Single-instance output format:
Diagnosis conclusion: instance i-bp1xxxxxxxxx (cn-shanghai) — 2 anomalies found (1 from Console Diagnosis, 1 from Cloud Assistant Diagnosis)
Console Diagnosis (1 anomaly):
[1] GuestOS.GPU.DriverNotInstalled
Severity: Warn
Description: Detect GPU driver installation status
Action: User needs to install driver
Cloud Assistant Diagnosis (1 anomaly):
[1] Device Driver Install Check - Failed
Description: Detect GPU driver installation status
Action: User needs to install driver
Recommendations:
- Install the matching version of the NVIDIA GPU driver
- Installation guide: https://help.aliyun.com/zh/egs/install-a-gpu-driver-on-a-gpu-accelerated-compute-optimized-linux-instance
Batch output format (Cloud Assistant Diagnosis only):
Diagnosis conclusion: cn-shanghai — 5 instances in total: 2 abnormal, 3 normal
========== ABNORMAL INSTANCES (2) ==========
--- Instance: i-bp1aaaaaaaaaa ---
[1] GuestOS.GPU.DriverNotInstalled — Detect GPU driver installation status. Action: install driver
[2] GuestOS.GPU.NVXidError — Detect GPU Xid error anomalies. Action: restart instance based on the XID error
Recommendations: Install NVIDIA driver; Restart instance to clear Xid errors
... (repeat per abnormal instance)
========== NORMAL INSTANCES (3) ==========
i-bp3ccccccccccc, i-bp4ddddddddddd, i-bp5eeeeeeeeeee — No anomalies detected
Special Reminder: When the exception handling measure is "O&M notification will be sent to user", append the following reminder:
⚠️ Important Reminder:
- Alibaba Cloud will send you O&M event notifications
- Please go to the ECS console to view event details
- Pay attention to whether you receive O&M events and handle them as required
Edge Case Handling
- Instance does not exist: CLI will return an error, capture and inform the user that the instance ID may be incorrect
- Region error: Prompt user to confirm the region where the instance is located
- Non-GPU specification: If the instance is not a GPU specification, diagnosis may have no results, prompt user to confirm instance type
- Insufficient permissions: If permission error is returned, prompt user to check AccessKey permissions
- Network timeout: Set command execution timeout (recommended 30 seconds), retry after timeout or prompt user to check network
- Cloud Assistant not available: If RunCommand returns an error indicating the cloud assistant is not installed or the instance is not running, inform the user: "Cloud Assistant Diagnosis is unavailable on instance ${INSTANCE_ID}. Please confirm the instance is in the Running state and the Cloud Assistant Agent is installed. Cloud Assistant Diagnosis for this instance has been skipped."
- Partial failure in batch: If some instances fail in Cloud Assistant Diagnosis (e.g., cloud assistant unavailable), continue with the remaining instances and report failures separately
- Scheduled: pure tag scheduling rejected: If
invoke-command returns MissingParam.InstanceId when only --tag is passed, resolve the tag into instance IDs via describe-instances first, then create the schedule with explicit --instance-id flags
- Scheduled: command name conflict: If
create-command fails due to a duplicate name, append a distinguishing suffix to gpu_diagnosis_<date> and retry
- Scheduled: user requests deletion: If the user asks to delete a scheduled task or command, do NOT delete any resource yourself per the "Resource deletion is STRICTLY FORBIDDEN" constraint. Instead: 1) execute
stop-invocation (with the task's InvokeId) to stop future scheduled runs — stopping is permitted and is the agent's way to halt the task; 2) tell the user that deletion must be done manually in the console: ECS console → Cloud Assistant → Command Execution Results → Scheduled Executions, locate the scheduled task and delete it there
Example Workflow
Single Instance:
User: Help me diagnose this GPU server i-bp1xxxxxxxxx
Agent:
1. Check CLI is installed
2. Ask for region (user did not provide)
3. User replies: cn-shanghai
4. Check instance OS type is Linux
5. [Console Diagnosis] Execute CreateDiagnosticReport, get ReportId: dr-xxxxxxxx
[Cloud Assistant Diagnosis] Execute RunCommand, get InvokeId: t-xxxxxxxx (started simultaneously)
6. Poll both DescribeDiagnosticReports and DescribeInvocationResults
7. Cloud Assistant Diagnosis finishes first — parse Base64 output (even if status is Failed/ExitCodeNonzero, decode Output)
8. Console Diagnosis finishes — parse Issues
9. Per Output Description, output BY METHOD DIMENSION — a Console Diagnosis section listing its anomaly items, then a Cloud Assistant Diagnosis section listing its anomaly items (NEVER merge the two methods' anomalies into one item), output to user
Scheduled Diagnosis (tag filtering):
User: Create a scheduled GPU diagnosis task for instances tagged gpu=1, running every day at 10:00
Agent:
1. Check CLI, validate region; resolve tag gpu=1 via describe-instances → Linux instances matched
2. Convert "every day at 10:00" to Cron: 0 0 10 * * ?
3. CreateCommand with the fixed script (Base64 literal, same as Cloud Assistant Diagnosis step 1), name gpu_diagnosis_<today>
4. InvokeCommand with repeated --instance-id flags + --frequency '0 0 10 * * ?'
5. Verify via describe-invocations (RepeatMode=Period, Scheduled); report IDs + result-query command
6. NEVER delete the command or task yourself; if later asked to delete, run stop-invocation and direct the user to delete it in the ECS console (Cloud Assistant → Command Execution Results → Scheduled Executions)
Batch Diagnosis:
User: Batch diagnose these GPU instances: i-bp1aaa, i-bp2bbb, i-bp3ccc, region cn-shanghai
Agent:
1. Check CLI; validate all instance IDs and region
2. Query all instances' OS type in one call, filter to Linux only
3. [Cloud Assistant Diagnosis ONLY] One RunCommand call with all instance IDs → 1 InvokeId
4. Poll DescribeInvocationResults; output each instance's result as soon as it is ready
5. Final summary: list normal instance IDs