spectredeath/skill-flywheel/skill-hardware-model-selector
Claude Code and agent skill by spectredeath/skill-flywheel
- community
Checking that the SKILL.md file is still in the repository.
Install
npx vskill@latest install spectredeath/skill-flywheel/skill-hardware-model-selectorSKILL.md
The same SKILL.md was published earlier as spectredeath/skill-flywheel/skill-hardware-model-selector.
origin: manual triggers:
- agent
- ai
- development
quality: applied_count: 0 success_count: 0 completion_rate: 0.0 token_savings_avg: 0.0 created_at: "2026-03-24T10:00:00Z" updated_at: "2026-03-24T10:00:00Z"
Hardware Model Selector
Description
This skill provides a standardized framework for hardware-aware model routing by mapping system-level hardware characteristics (e.g., VRAM, CUDA cores) to model requirements.
Purpose
Enables agents to autonomously determine the optimal LLM configuration based on available hardware profiles and resource constraints.
Capabilities
- Hardware discovery and VRAM telemetry
- Profile generation for heterogeneous clusters
- Quantization-aware model mapping (4-bit/AWQ vs FP16)
- Safety margin calculations for KV cache headroom
Usage Examples
- "Select a model for a laptop with 8GB VRAM" -> Result: "Llama-3-8B-Q4"
- "Select a model for an A100 cluster for 405B execution" -> Result: "Llama-3.1-405B-FP8"
Input Format
JSON object containing hardware_stats (VRAM, compute_cap) or a string describing the device.
Output Format
Recommended model ID, quantization level, and execution strategy.
Configuration Options
min_headroom_gb: Default 2.0preferred_quantization: Default AWQ
Constraints
- ALWAYS check available VRAM before model selection.
- NEVER exceed 90% of available GPU VRAM.
- MUST provide a fallback model in every selection plan.
Examples
Example 1: High-End Consumer
Input: RTX 4090 (24GB) Output: Llama-3.1-70B-AWQ
Error Handling
- Case: OOM Imminent -> Action: Downscale to 8B model.
- Case: Driver Mismatch -> Action: Force CPU fallback or basic quantization.
Performance Optimization
- Cache hardware profiles for 300s to avoid redundant
nvidia-smicalls. - Use pre-calculated VRAM tables for all registered models.
Integration Examples
- Integrated into
AgentOrchestratormission planning lifecycle. - Called via MCP tool
model_select.
Best Practices
- Always prioritize quantization for consumer hardware.
- Maintain a 2GB VRAM "System Buffer" at all times.
Troubleshooting
- Symptom: Model load failure -> Check:
nvidia-smifor hidden VRAM consumers. - Symptom: Slow TTFT -> Check: Bus bandwidth vs model size.
Monitoring and Metrics
- Track VRAM peaks and model load success rates.
- Monitor
hardware_selection_latency(Target < 10ms).
Dependencies
performance_monitormodel_registrynvml-python
Version History
- 1.0.0: Initial release
- 4.2.0: Synchronized with Skill Flywheel industrial standard
License
MIT License
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingdomains/model_orchestration/SKILL.hardware-model-selector/SKILL.md
Popularity
- Repository stars
- 1
- Repository forks
- 1
- 7-day trend score
- 8
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/spectredeath/skill-flywheel/skill-hardware-model-selector)