qzcli

Manage GPU compute jobs on the Qizhi platform using a kubectl-style CLI.

Updated Jul 9, 2026
One-click install
npx skills add https://github.com/Lingjie-wang/autoRL --skill qzcli-lingjie-wang
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: qzcli
Source: https://github.com/Lingjie-wang/autoRL/tree/main/Auto-claude-code-research-in-sleep/skills/qzcli
Command: npx skills add https://github.com/Lingjie-wang/autoRL --skill qzcli-lingjie-wang

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires rich, requests, prompt_toolkit, mcp.

What problem does it solve? Submitting, monitoring, and stopping distributed GPU training jobs on the Qizhi (启智) platform normally requires manual web console work; this Skill wraps the qzcli CLI so jobs, workspaces, and compute groups can be managed from the command line or an AI agent. ## Core Features & Use Cases - Job Lifecycle Management: Login, discover workspaces and compute groups, check available GPU nodes, submit jobs interactively or non-interactively, and stop or watch running jobs. - Batch & Matrix Submission: Submit Cartesian-product job sweeps from a JSON config with name and command templates, plus Slurm-based HPC/CPU job support. - MCP Integration: Expose qzcli as an MCP tool so Claude Code or Codex can manage jobs directly. - Use Case: A researcher needs to evaluate a model at four checkpoints across two steps. They run qzcli res -u to refresh the workspace cache, then qzcli batch batch_config.json --dry-run to preview eight jobs before submitting them with a 3-second delay between submissions. ## Quick Start Ask the assistant to log in to the Qizhi platform with qzcli, refresh the workspace resource cache, and list currently running jobs.

Frequently Asked Questions about qzcli

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I submit a GPU training job on the Qizhi platform?

Use qzcli create with a job name, command, workspace, compute group, and Docker image. Run qzcli create -i for an interactive TUI that walks through workspace, project, compute group, and spec selection, or pass IDs directly for scripted submission.

How do I submit multiple jobs with different parameters in batch?

Use qzcli batch with a JSON config defining defaults, a matrix of parameter values, and name/command templates. Matrix keys are Cartesian-producted, and you can preview everything with --dry-run before submitting.

Can I use qzcli with Claude Code or Codex as an MCP tool?

Yes, qzcli ships an MCP server. After installing with pip install -e ., register it using claude mcp add qzcli -- qzcli-mcp for Claude Code or codex mcp add qzcli -- qzcli-mcp for Codex.

Why does qzcli say my workspace was not found?

The error '未找到名称为 xxx 的工作空间' means the local resource cache is stale. Run qzcli res -u to refresh the workspace and compute group cache, then retry the command.

How does qzcli handle authentication and expired cookies?

Credentials are read from CLI args, stdin, environment variables, a .env file, or ~/.qzcli/config.json in that priority order. The qzcli avail command auto-refreshes an expired cookie when credentials are configured; otherwise re-run qzcli login.

Does qzcli support CPU-only or Slurm HPC jobs?

Yes, the qzcli hpc subcommand submits Slurm-based CPU jobs with parameters for CPU count, memory in GiB, instance count, predefined quota ID, and an entrypoint command.