Parallel Device Config

Parallel Anything is the orchestra; these are the musicians. Parallel Device Config is the one-per-device building block: each node adds a single GPU (or CPU) to a DEVICE_CHAIN, and you wire several together to tell Parallel Anything what hardware you've got and how to split work across it.

What it is

One node, one device, one percentage. Chain them by feeding the first node's device_chain output into the next node's previous_devices input, and so on down the line. The final device_chain output goes into the Parallel Anything node. The mental model is simple: the chain is just a list of {device, percentage} entries, and each Config node appends one.

The two inputs that matter

• device_id - a dropdown auto-populated from what's actually in your machine: cpu, cuda:0, cuda:1, and so on; mps on Apple Silicon; xpu on Intel; even DirectML devices if you have torch-directml. You can only pick devices that exist, which removes a whole class of typos.
• percentage - how much of the work this device does. The author's own tooltip is the best description: "Percentage of batch (or layers for batch=1) to process on this device." Batch > 1 means the batch gets split data-parallel style; batch == 1 means the transformer blocks get split pipeline-style instead.

Percentages are relative, not absolute - Parallel Anything normalizes them before splitting, so a 50/50, a 60/40, and a 30/20 all behave the way you'd expect.

Which device goes first

The first device in the chain is the lead device - the one results get gathered and concatenated on, and the one that runs the whole model if the batch is too small to split. Lead with your fastest card. If your GPUs are mismatched, the README's advice holds: put the bigger percentage on the faster card rather than expecting them to keep pace with each other.

Install and gotchas

Same pack install as Parallel Anything:

cd ComfyUI/custom_nodes
git clone https://github.com/FearL0rd/ComfyUI-ParallelAnything.git

Then restart ComfyUI (or use ComfyUI Manager, search "ComfyUI-ParallelAnything"). No extra dependencies.

Two quick gotchas. First, this pack is pointless on a single-GPU box - you need at least two devices, and the README allows "1 GPU + CPU" only for testing. Second, the dropdown shows only detected devices: if cuda:2 doesn't appear, you don't have three GPUs, and there's no way to force it in.

For a fixed 2–4 GPU workstation, Parallel Device List gives you the same result in one node instead of a chain - four dropdowns beat four wired nodes when your hardware never changes.