Skip to main content
The Select Model Device node allows you to manually choose which device (CPU or a specific GPU) a diffusion model runs on. It can move a model to a different device, and it handles conflicts with other multi-GPU nodes automatically. Selecting "default" restores the original device chosen by the model loader.

​
Inputs

Parameter Details:
  • "default": Restores the load and offload devices assigned by the model loader, even after a previous Select Model Device call.
  • "cpu": Pins both the load and offload device to CPU.
  • "gpu:N": Pins the load device to the Nth available GPU (for example, "gpu:0" for the first GPU). The offload device is restored to the loader’s original choice.
Important Notes:
  • If the requested device does not exist on the current machine (for example, a workflow created on a 2-GPU machine is opened on a 1-GPU machine), the node passes the model through unchanged and logs a message instead of failing. Unknown gpu:N values are allowed during input validation so portable workflows do not error early.
  • If the model is already on the requested device, the node uses a fast path and does not reload the model.
  • When the requested device differs from the device the input model is already on, a fresh model is spawned using the loader’s reload factory so the new patcher owns independent weights on the new device.
  • If the model loader does not support multi-GPU (no reload factory), the node passes the model through unchanged and logs a warning.
  • When a MultiGPU CFG Split clone already occupies the selected device, that clone is pruned so two models are not bound to the same device.
  • When a non-default device is selected (CPU or GPU), the node also adjusts the model’s compute dtype to one supported by that device.
  • Placing this node after a node that has already consumed the model (for example, a KSampler) is not recommended, because any state changed by the prior node will be observed if the device matches the original.

​
Outputs

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub
Source fingerprint (SHA-256): d02a8bd9612861cf696f03969fe693088351de5a72ccbd4c1aed405b104eb71e