FearL0rd ComfyUI-ParallelAnything Tool for 2 or more GPUs
ComfyUI-ParallelAnything is a custom node suite for ComfyUI that enables true parallel processing across multiple GPUs. Unlike standard offloading methods that move a single model between devices, this tool creates independent model replicas on each selected GPU or CPU for simultaneous batch processing.
FearL0rd, the developer who built this solution for users with multi-GPU workstations who want to maximize their hardware utilization. The tool supports configurations ranging from two GPUs up to eight or more, with dynamic load balancing to distribute work efficiently across different hardware combinations.
Key functionality of ComfyUI-ParallelAnything
- True parallel execution with simultaneous forward passes across multiple GPUs.
- Chainable device nodes that let users u tp 2-8+ GPU setups.
- Automatic hardware detection that populates dropdown menus with available hardware.
- Dynamic load balancing with percentage-based batch splitting.
- Cross-platform support for Windows, Linux, and macOS with MPS backend.
Users with mixed GPU setups can benefit from the percentage-based workload distribution, which compensates when pairing faster cards with slower ones. For more hardware management in ComfyUI FearL0rd also released ComfyUI-AnyDeviceOffload.
Technical requirements and notable limitations
The developer notes that this approach uses model replication rather than model parallelism, meaning each GPU must hold a complete copy of the model. Users should expect VRAM usage to scale linearly with the number of GPUs. The documentation states:
'If you use 2 GPUs, you will use 2 times the VRAM (one copy per card)'
Batch size plays a critical role in performance. Parallelism only activates when the batch size exceeds the number of devices in the chain. The tool handles inference only and does not support training or fine-tuning workflows. Users mixing different GPU architectures should adjust percentage weights to account for speed differences between cards.
