flow_low_gpu

No description ships with this node, but the field names point pretty clearly at VRAM management: reserved and a manual/auto mode toggle. Read alongside its anything → output pass-through pattern - a very common ComfyUI trick for forcing a node to execute at a specific point in the graph, since ComfyUI otherwise schedules nodes by data dependency rather than canvas position - the likely job here is to sit at a chosen spot in your workflow (say, right before a heavy sampler) and make sure some VRAM headroom is reserved or checked at that exact moment, then pass whatever's wired into it straight through unchanged.

How it works

anything accepts any type and output returns the same value unmodified - that pass-through is what lets you insert this node into an existing chain (between a loader and a sampler, for instance) purely to force execution ordering, without altering the data flowing through it. reserved (a float from 0 to 24, presumably gigabytes) is the amount of VRAM you want held back rather than used, and mode switches between auto (the default - let the node figure out how much to reserve) and manual (use the reserved value directly).

Since none of this is spelled out in the pack's README, treat it as an informed read of the parameters rather than confirmed behavior - worth watching your GPU memory usage (nvidia-smi or ComfyUI's own memory logging) with and without this node in a graph to confirm what it's actually doing before relying on it to prevent an OOM.

Inputs and output

• anything (required, wildcard) - passed straight through.
• reserved (FLOAT, default 0.6, range 0–24) - VRAM to hold back, in manual mode.
• mode (enum: manual / auto, default auto) - how the reservation amount is decided.
• output (wildcard) - the pass-through of anything. This node is also flagged as an output node, so it's valid to leave it as a dead end too.

Installing it

Search ComfyUI-Apt_Preset in ComfyUI Manager and install, or manually: cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git, then restart. The README's own install step is "double-click install.bat" - Windows-first, with no requirements.txt visible elsewhere, so on Linux/Mac open that file to see what it installs and mirror it manually in your ComfyUI venv. This node is pure control flow and needs none of the pack's optional extras (GGUF, Advanced-ControlNet, nunchaku, downloaded models).

Common issues

Since the exact mechanism isn't documented, don't treat this as a guaranteed fix for out-of-memory errors - it's more likely a scheduling/checkpoint aid than an actual memory-limiting mechanism (ComfyUI itself, not a custom node, ultimately controls what gets loaded onto the GPU and when). If you're chasing a real OOM, the standard fixes - smaller batch sizes, lower resolution, offloading models, using --lowvram/--novram launch flags - are more reliably effective than any single node.

Placement in the graph likely matters here given the pass-through-for-ordering pattern - putting it in the wrong spot (after the memory-heavy node instead of before it) would defeat the purpose. Test with something you can watch VRAM usage against.

And the pack-wide note: this is one node in a large, actively developed package, and an import error elsewhere in the pack can take this node down too. If it's missing after install, check the ComfyUI console log at boot for the real traceback.