Coordinate GPU VRAM usage across Ollama, Whisper, and ComfyUI with OOM retry and auto-unload.
npx skills add https://github.com/lawless-m/Gwen --skill vram-gpu-oom
Please help me install this Agent Skill. Skill: Vram-GPU-OOM Source: https://github.com/lawless-m/Gwen/tree/main/.claude/skills/Vram-GPU-OOM-memory-management Command: npx skills add https://github.com/lawless-m/Gwen --skill vram-gpu-oom
This Skill provides GPU VRAM sharing and OOM retry patterns to coordinate memory across Ollama, Whisper, and ComfyUI, with auto-unload and signaling endpoints.
Implement OOM retry logic and configure auto-unload; test with smaller models before scaling up.
High-intent search queries and answers about installing and using this skill.
Coordinate GPU VRAM usage across Ollama, Whisper, and ComfyUI by implementing OOM retry logic with configurable delays and automatic cache cleanup. The Skill enables cross-service signaling so services can request unloads when memory pressure rises, preventing crashes.
Yes. The Skill provides auto-unload on idle to release VRAM without manual intervention. Configure idle thresholds and the system will unload models from memory, freeing capacity for other services while maintaining quick restart capability.
Use cross-service OOM retry with delays and retry counts. When one service hits OOM, the Skill signals others to unload, waits a configurable interval, then retries the failed operation. This pattern prevents cascading failures across your AI service stack.
The Skill includes optional signaling endpoints to request unloads and report service status. Use these to integrate with your workflow, or rely on built-in auto-unload and retry without additional API calls if your use case permits.
Coordination is designed for Ollama, Whisper, and ComfyUI. The Skill applies OOM detection, retry logic, and auto-unload patterns to these three services, with signaling endpoints for cross-service communication.
Snowflake-LabsA concise summary of your roleplay character: name,</name> personality, and/or behavior, formatted as a single string sentence. Typically不超过50字，此处省略
celticht32Monitor Couchbase cluster health and configure Prometheus and Grafana alerts.
freedomw1987Manage s6-overlay supervision trees in Hermes Agent Docker containers.
ovachieverQuantize LLMs to 8-bit or 4-bit precision using BitsAndBytes.
