菜单
Profile actual GPU memory usage per model and compare to declared memory_gb in models.yaml. Identifies waste, tight allocations, and optimization opportunities. Use when asked to "profile memory", "check memory usage", "optimize memory", or "how much memory is each model using".
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/kshetrajna12/sparkstation --skill sparkstation-profile-memory命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Safe deployment operations for Sparkstation. Profile switching with drain/health-check, rolling model updates, Docker image rebuilds, full stop/start, and rollback. Use when asked to "deploy", "switch profile", "restart sparkstation", "update model", "rebuild container", or "rollback".
Benchmark inference throughput and latency for Sparkstation models. Measures TTFT, tokens/sec, p50/p95/p99 latency, and embedding throughput. Use when asked to "benchmark", "measure latency", "test throughput", or "profile inference performance".
Validate models.yaml profiles for correctness, consistency, and resource feasibility. Checks memory limits, alias collisions, required args, backend compatibility. Use when asked to "lint config", "validate models.yaml", "check config", or before deploying a new profile.
基于 SOC 职业分类
正在显示 SKILL.md
| name | sparkstation-profile-memory |
| description | Profile actual GPU memory usage per model and compare to declared memory_gb in models.yaml. Identifies waste, tight allocations, and optimization opportunities. Use when asked to "profile memory", "check memory usage", "optimize memory", or "how much memory is each model using". |
Measures actual memory usage per model and compares to declared memory_gb in models.yaml.
# Profile all running models
python3 .pi/skills/sparkstation-profile-memory/scripts/profile_memory.py
# JSON output
python3 .pi/skills/sparkstation-profile-memory/scripts/profile_memory.py --json
# Include models.yaml analysis (declared vs actual)
python3 .pi/skills/sparkstation-profile-memory/scripts/profile_memory.py --compare
For each running model:
docker stats (actual process memory)models.yaml profile config═══ SPARKSTATION MEMORY PROFILE ═══
MODEL MEMORY USAGE
qwen3-vl-30b declared: 55.0 GB actual: 48.2 GB delta: -6.8 GB ⚠️ over-allocated
bge-m3 declared: 2.5 GB actual: 2.1 GB delta: -0.4 GB ✅ OK
clip-vit declared: 5.0 GB actual: 1.8 GB delta: -3.2 GB ⚠️ over-allocated
species-detect declared: 5.0 GB actual: 1.7 GB delta: -3.3 GB ⚠️ over-allocated
SUMMARY
Total declared: 67.5 GB
Total actual: 53.8 GB
Potential savings: 13.7 GB (20.3%)