A VRAM scheduler for providing local AI companionship to AstrBot and DeepSeek Harness: On the same 12 GB card, it stops unnecessary local LLM, IndexTTS, or ComfyUI processes based on the task, starts the required ones, and restores them after use.
Stars 28
Reasons
- Addresses the issue where VRAM overflow on Windows does not throw an error but instead pages VRAM to system memory, causing extreme slowdowns; the author claims this reduced voice diffusion time from 30 seconds per step back to 0.12 seconds on their machine.
- Only stops and starts processes listed in your profile (process name + command line + port whitelist); it waits for services to be truly ready and for the ComfyUI queue to empty before switching; the DSH side provides vram_status / vram_acquire / vram_release tools, along with an MCP entry point.
Usage
- git clone https://github.com/xiaoxiao150/vram-switcher && cd vram-switcher, then run python configure.py --root D:/AI to automatically detect running services and generate profile.local.json (add --dry-run to preview without writing).
- AstrBot users can search for VRAM Switcher in the plugin marketplace; voice integration requires your TTS plugin to write and delete files according to the tts_busy.flag / tts_ready.flag protocol.
Risks
- The relationship with DeepSeek is limited to supporting DSH as a host; the local dialogue model in the author's example is not DeepSeek. The use case is narrow, having been production-verified only on NVIDIA GPUs on Windows.
- It terminates and starts local processes according to the configuration; an incorrect profile may stop services that should not be stopped.