GPU showing full but Task Manager lying to you?

Microsoft's Scott Hanselman just dropped gputaskman - basically htop for your GPU.

Shows you:
→ Real VRAM usage per process
→ Which model files are actually loaded
→ For local LLMs: model weights vs context (KV cache) breakdown
→ What's secretly eating your system RAM instead of GPU memory

Works on Windows (NVIDIA/AMD) + Linux/WSL (NVIDIA). Portable download, zero install.

If you're running local AI models and constantly hitting "out of memory" errors while Task Manager gaslights you, this is it.