GPU disappears during video generation when splitting UNet/CLIP/VAE across multiple GPUs with MiniMax H3 (Ubuntu + 4x RTX 3090)

[segatecm]

Description

When using ComfyUI-MultiGPU to distribute different model components (UNet, CLIP, VAE) to separate GPUs while running MiniMax H3 video generation, one of the GPUs (the one assigned to UNet) disappears / becomes invisible during the generation process.

The system can no longer see that GPU (it no longer shows up in nvidia-smi), and the generation fails.

Steps to Reproduce

1. Install ComfyUI-MultiGPU (latest version).
2. Load a MiniMax H3 workflow (text-to-video or image-to-video).
3. Use MultiGPU loader nodes to assign:
   • UNet → one GPU (e.g. cuda:0)
   • CLIP → another GPU (e.g. cuda:1)
   • VAE → another GPU (e.g. cuda:2)
4. Start video generation.
5. During the generation process, the GPU assigned to UNet disappears.

Expected Behavior

All GPUs should remain visible and stable throughout the entire generation process.

Actual Behavior

The GPU that has the UNet loaded becomes invisible / lost during generation (no longer appears in nvidia-smi). Generation then fails.

Environment

• OS: Ubuntu
• GPUs: 4 × NVIDIA RTX 3090

Additional Information

• This only happens when splitting UNet / CLIP / VAE across different GPUs.
• Happens during the video generation process (not immediately at model loading).
• System has 4× 3090, so there should be enough VRAM.