Dynamic VRAM is a ComfyUI model-loading mode. The flag that turns it off, --disable-dynamic-vram, is described in the CLI as "disable dynamic VRAM and use estimate based model loading", so dynamic VRAM is the alternative to loading models by estimating up front what fits.
The flags
From cli_args.py:
--enable-dynamic-vram: enable it on systems where it is not enabled by default.--disable-dynamic-vram: disable it and use estimate based model loading.--vram-headroom GB: extra VRAM for Dynamic VRAM to keep free above the default.--fast-disk: force disk-backed dynamic loading and offload, which can be faster on fast NVMe disks.
When it is off
In the current source, enables_dynamic_vram() returns true unless --disable-dynamic-vram is set or you use --highvram, --gpu-only, --novram or --cpu. Passing --enable-dynamic-vram forces it on.
Why it matters for other flags
The help text for --lowvram says it does nothing when dynamic VRAM is enabled, so on a system using it, --lowvram is not the lever. See ComfyUI low VRAM mode. --reserve-vram is used as its headroom: what --reserve-vram does.
If a model fails to load, see how to fix ComfyUI out of memory errors.
Sources
- ComfyUI cli_args.py (dynamic VRAM flags, enables_dynamic_vram). github.com
- ComfyUI main.py. github.com