CheckpointLoaderSimpleDisTorch2MultiGPU

This is the DisTorch2 version of the classic "load everything" node. Point it at a single-file checkpoint - SD1.5, SDXL, anything still packed as one .safetensors bundling the model, CLIP, and VAE together - and it hands back all three, just like core ComfyUI's CheckpointLoaderSimple. The file picker's own tooltip says it plainly: "The name of the checkpoint (model) to load." What's different from the plain multigpu version is the memory handling underneath.

How it works

There's one set of DisTorch2 controls here - compute_device, virtual_vram_gb, donor_device, expert_mode_allocations, eject_models - not three separate ones per component. In practice, that split targets the diffusion weights: the largest chunk of the file by a wide margin. CLIP and VAE ride along without their own independent offload knobs. If you need to tune all three components separately, that's what CheckpointLoaderAdvancedDisTorch2MultiGPU is for - this node deliberately keeps a smaller interface.

The payoff: someone running an older-style, one-file SD1.5 or SDXL checkpoint gets the same "make VRAM feel bigger" trick that Flux and Wan users get from UNETLoaderDisTorch2MultiGPU, without needing to go find separate UNet/CLIP/VAE files first.

The inputs and outputs that matter

• ckpt_name - the checkpoint file.
• compute_device, virtual_vram_gb (default 4), donor_device, expert_mode_allocations - same DisTorch2 pattern as the standalone UNet loader; see UNETLoaderDisTorch2MultiGPU for the full tuning walkthrough (short version: don't max the slider, only offload what you need, watch the console's memory summary at load time).
• eject_models (default true) - fully drops the model when done, which makes sense here since a checkpoint-style node is the one you're most likely to swap out entirely between sessions.
• Outputs, carrying over the core ComfyUI tooltips: MODEL ("the model used for denoising latents"), CLIP ("the CLIP model used for encoding text prompts"), VAE ("the VAE model used for encoding and decoding images to and from latent space").

How to install it

ComfyUI Manager: search ComfyUI-MultiGPU, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/pollockjj/ComfyUI-MultiGPU

then restart. No extra dependencies.

Common issues & troubleshooting

Overshooting virtual_vram_gb. Same tuning story as everywhere else in the DisTorch2 family - offloading past what your run actually needs just adds PCIe overhead for nothing. GPU memory bandwidth genuinely outruns the PCIe bus by roughly an order of magnitude, so this cost is real, not theoretical.

Wrong node for a split checkpoint. If you're actually running a modern setup with separate UNet + CLIP + VAE files, this node won't help you - it only reads single-file bundles. Reach for UNETLoaderDisTorch2MultiGPU, CLIPLoaderDisTorch2MultiGPU, and VAELoaderMultiGPU instead, each with its own independent controls.

eject_models surprises. It defaults to true here - full unload between runs - which differs from the CLIP-family DisTorch2 loaders that default to false and stay cached. If you're swapping between two checkpoints repeatedly and it feels slower than expected, that eager unload is why.