sparkstation-profile-memory

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".

Sparkstation Memory Profiler

Measures actual memory usage per model and compares to declared memory_gb in models.yaml.

Usage

# 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

What It Measures

For each running model:

• Docker stats memory — RSS from docker stats (actual process memory)
• Declared memory_gb — from models.yaml profile config
• Delta — over/under-allocation
• Recommendations — tighten or loosen allocations

Output

═══ 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%)