Comfy Team Revolutionizes AI Memory with ComfyUI Dynamic VRAM

ComfyUI Dynamic VRAM is a new memory optimization system designed to help users run large AI models on hardware with limited memory. It introduces a custom PyTorch VRAM allocator that manages model weights more efficiently, reducing crashes and improving performance. The system is now available in the stable release of ComfyUI for Nvidia hardware on Windows and Linux.

This update addresses the growing issue of high RAM prices by making better use of existing hardware resources. Developed by the Comfy team, it fundamentally changes how the application handles memory allocation during inference. The goal is to make large open-source models more accessible to users who cannot afford expensive hardware upgrades.

Smarter memory management

• Reduces system RAM usage significantly during complex workflows.
• Eliminates Out-Of-Memory crashes caused by insufficient weight offloading.
• Speeds up initial model loading and LoRA applications.
• Prevents slow page file usage when running models that exceed physical RAM.
• Increases GPU VRAM utilization for faster processing.

Users working with limited hardware configurations can benefit from this system, particularly those running multiple models or video generation workflows. The technology allows weights to stay in VRAM for speed while freeing them instantly when memory pressure occurs, creating a smoother experience without manual quota management.

Technical implementation details

The system uses a Virtual Base Address Register (VBAR) that consumes no physical VRAM when created, instead using GPU virtual address space. When a model layer needs a specific weight, a custom fault() API allocates the memory at the exact moment it is required.

The team notes that users may see higher VRAM usage in Task Manager, which is expected behavior. As the developers explain,

'these cached weights will never be pushed to your page file.'

Future updates plan to add AMD support and further reduce RAM footprint through experimental features.

Read more about ComfyUI Dynamic VRAM on their official blog.