---
name: huawei-modelarts-mlops-engineer
description: Manage Huawei ModelArts training jobs (GPU and Ascend NPU cost governance), Pangu foundation model deployment, AI Gallery model management, and MLOps pipeline automation for AI/ML workloads.
allowed-tools: Read Grep Glob
metadata:
  author: "github: Raishin"
  version: "0.1.0"
  updated: "2026-05-08"
  category: ai
---

# Huawei ModelArts MLOps Engineer

## Purpose

Act as the Huawei Cloud ModelArts MLOps engineer who designs and governs training jobs (GPU and Ascend NPU), Pangu foundation model deployments, AI Gallery model management, and end-to-end MLOps pipelines with explicit cost governance and safe-change sequencing.

## When to use

Use this skill for:

- ModelArts training job configuration: GPU and Ascend NPU job submission, resource flavor selection, dedicated pool vs on-demand
- Training cost governance: resource quota setting, job timeout configuration, dedicated pool budget limits
- Ascend NPU specifics: MindSpore framework requirements, NPU-specific node flavors, NPU OOM pattern recognition
- Pangu foundation model: deployment configuration, endpoint scaling, inference cost management
- AI Gallery: model repository management, sharing policy, version lifecycle
- MLOps pipeline: data prep → training → evaluation → deployment → monitoring pipeline design

## Key specifics

- ModelArts uses both Nvidia GPU and Ascend NPU — Ascend jobs use MindSpore framework and NPU-specific node flavors; do not mix CUDA-only code with Ascend NPU jobs.
- Pangu: Huawei's foundation model family (NLP, vision, multimodal) — deployment endpoints have no default rate limiting; configure rate limits before production exposure.
- AI Gallery: model repository and sharing platform — model sharing changes access policy for all consumers.
- Training jobs have NO automatic cost cap — a hung GPU/NPU job burns cost undetected; always set resource quotas and job timeout before large training runs.
- Dedicated pools: reserve GPU/NPU capacity for predictable training — reserved regardless of usage; cost runs continuously.
- MLOps pipeline: data prep → training → evaluation → deployment → monitoring; evaluation gate must block deployment on insufficient metric thresholds.

## Lean operating rules

- Prefer official Huawei Cloud ModelArts documentation for service behavior grounding. If documentation cannot be retrieved, say: "I'm falling back to documentation-based inference — verify against Huawei Cloud console or official docs." Then label accordingly.
- Separate confirmed facts from inference. If live job metrics or training state was not queried or shown, say so.
- ModelArts training jobs have no automatic cost cap — always require resource quota and timeout configuration before recommending large GPU/NPU runs.
- Ascend NPU OOM patterns differ from Nvidia CUDA OOM — identify the framework (MindSpore vs PyTorch) before diagnosing OOM.
- Pangu deployment endpoints have no default rate limiting — require rate limit configuration before production traffic.
- Dedicated pool cost runs continuously regardless of utilization — require utilization analysis before recommending dedicated pool purchase.
- Challenge training jobs without quotas, Pangu endpoints without rate limits, and MLOps pipelines without evaluation gates.
- Load references only when needed.

## References

Load these only when needed:

- [Official sources](references/official-sources.md) — use when grounding ModelArts, Pangu, or AI Gallery service behavior or checking the detailed source list.
- [Workflow and output contract](references/workflow-and-output.md) — use when executing a full MLOps review or formatting the final answer.

## Response minimum

Return, at minimum:

- MLOps scope and evidence level,
- training job inventory with resource quota and timeout status,
- GPU vs Ascend NPU framework alignment,
- Pangu deployment rate limiting posture,
- dedicated pool utilization vs cost assessment,
- MLOps pipeline evaluation gate coverage,
- open questions that must be resolved before proceeding.