Two GPUs, one active device: distinguish selection, concurrency and memory pooling

Define which multi-GPU behavior you need

We distinguish choosing a device for one task, running separate tasks on separate devices, and splitting one model across devices. The presence of two cards does not establish any of those behaviors automatically.

PyTorch documents CUDA_VISIBLE_DEVICES as a visibility setting. ComfyUI exposes system information through its API. 1 2 Neither is evidence that a workflow pools VRAM.

Our suggested record includes device names and stable identifiers, launch settings and backend logs. A screenshot showing only “GPU 0” is not enough to establish physical-device identity across differently configured processes.

First observe a small task. Change device selection only through a supported launcher or backend setting, restart, then compare actual device information. Do not assume every custom node follows the same selection path.

For a model larger than one device's memory, look for explicit model/backend support for partitioning or offloading. Simply listing more visible GPUs is not that implementation. Record device use, concurrency and outputs separately, and leave unmeasured resource figures blank.

Related checks

• Optional backend warnings: decide what blocks your workflow
• Keep a reproducible ComfyUI repair record without exposing secrets

Source-reviewed diagnostic guidance; no local multi-GPU test or measured VRAM result. Indexing remains withheld by site-wide gates.